feat: add public C++ operator API

[voltjia]

Summary

• Add the unified public entry header include/infini/ops.h for C++ consumers.
• Generate public one-shot operator declarations under infini::ops::functional in generated/include/infini/functional_ops.h.
• Compile generated functional operator definitions into libinfiniops.
• Route pybind one-shot operator calls through infini::ops::functional::<OpName>.
• Install public headers, libinfiniops, pkg-config metadata, and CMake package metadata.

Motivation

This is the first PR in the public operator API split. It adds the C++ source/link convenience layer first so the later stable C ABI can be implemented as a thin adapter over public C++ operator entrypoints instead of including backend kernel headers directly.

Closes #N/A

Type of Change

• feat - new feature / new operator / new platform
• N/A - fix - bug fix
• N/A - perf - performance improvement (no behavioral change)
• N/A - refactor - code restructuring without behavior change
• N/A - test - adding or fixing tests only
• N/A - docs - documentation only
• build / ci - build system or CI configuration
• N/A - chore - tooling, formatting, or other non-code changes
• N/A - Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Test Results on Supported Platforms

All platform tests below were run remotely with unqualified pytest and the platform bound to physical card 6.

Platform	Built	pytest Result	Notes / Hardware
NVIDIA	Yes	9206 passed, 8665 skipped, 81 warnings in 362.25s	ssh nvidia, CUDA_VISIBLE_DEVICES=0 inside Docker with --gpus device=6.
Iluvatar	Yes	7704 passed, 8649 skipped, 81 warnings in 575.63s	ssh iluvatar, CUDA_VISIBLE_DEVICES=6; CoreX compiler bin mounted at /usr/local/corex-4.3.0.20250624/bin.
MetaX	Yes	8698 passed, 7655 skipped, 81 warnings in 401.38s	ssh metax, CUDA_VISIBLE_DEVICES=6, MACA_VISIBLE_DEVICES=6.
Cambricon	Yes	5899 passed, 10070 skipped, 172 warnings in 991.75s	ssh cambricon, MLU_VISIBLE_DEVICES=6.
Moore	Yes	8471 passed, 7900 skipped, 99 warnings in 584.02s	ssh moore, MUSA_VISIBLE_DEVICES=6; no _musa_smoke_passes guard.
Ascend	Yes	7405 passed, 8906 skipped, 98 warnings in 552.77s	ssh ascend, ASCEND_RT_VISIBLE_DEVICES=6; pytest inner return code was 0.

Full `pytest` output (optional)

# NVIDIA
========= 9206 passed, 8665 skipped, 81 warnings in 362.25s (0:06:02) ==========

# Iluvatar
========= 7704 passed, 8649 skipped, 81 warnings in 575.63s (0:09:35) ==========

# MetaX
========= 8698 passed, 7655 skipped, 81 warnings in 401.38s (0:06:41) ==========

# Cambricon
======== 5899 passed, 10070 skipped, 172 warnings in 991.75s (0:16:31) =========

# Moore
========= 8471 passed, 7900 skipped, 99 warnings in 584.02s (0:09:44) ==========

# Ascend
========= 7405 passed, 8906 skipped, 98 warnings in 552.77s (0:09:12) ==========
PYTEST_INNER_RC=0

Additional checks:

git diff --check
python3 -m py_compile tests/utils.py scripts/generate_wrappers.py
ruff format --check tests/utils.py scripts/generate_wrappers.py
ruff check tests/utils.py scripts/generate_wrappers.py

Benchmark / Performance Impact

N/A. This PR adds public API wrappers and build/install plumbing. It does not change operator kernels or dispatch selection.

Notes for Reviewers

• The branch was rebased onto origin/master at 1400daff; current head is 4f451af0.
• The C++ API is intentionally a source/link convenience layer, not a stable C++ ABI commitment.
• Public one-shot calls live in infini::ops::functional to avoid colliding with existing operator classes such as infini::ops::Add.
• Backend implementation headers remain in generated source files; the public header does not include backend kernel headers.
• The wrapper generator now has a constrained fallback parser for the current base operator headers when Python clang.cindex is unavailable in direct CMake builds.
• The WITH_TORCH generated include directory is build-interface-only, which fixes the CMake install/export failure seen in wheel builds.
• _musa_smoke_passes was removed. The low Moore count was caused by filtering out musa; selecting physical card 6 with MUSA_VISIBLE_DEVICES=6 makes the normal torch.musa.is_available() path work.
• Ascend is no longer CPU-only in this result; ASCEND_RT_VISIBLE_DEVICES=6 produced ('cpu', 'npu') device discovery before the full pytest run.
• This PR should land before the C ABI PR. The follow-up C ABI PR should target this branch.

Generated public C++ signature examples:

void infini::ops::functional::Add(const Handle& handle, const Config& config,
                                  const Tensor input, const Tensor other,
                                  Tensor out);

void infini::ops::functional::Matmul(const Handle& handle, const Config& config,
                                     const Tensor a, const Tensor b, Tensor c,
                                     bool trans_a, bool trans_b);

void infini::ops::functional::Linear(const Handle& handle, const Config& config,
                                     const Tensor a, const Tensor b,
                                     std::optional<Tensor> bias, bool trans_a,
                                     bool trans_b, Tensor out);

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits: feat: add public C++ operator API.
• Branch name follows the expected <type>/... form: feat/cpp-operator-api.
• Commit messages follow Conventional Commits.
• Small PR is a single squashable change set.
• Branch was rebased onto current origin/master.
• No stray merge commits from master.
• No fixup! / squash! / wip commits remain.

Scope and Design

• Changes are minimal for the stated motivation.
• No dead code, commented-out blocks, debug prints, or ownerless TODOs were added.
• No unrelated formatting churn was added.
• Public API changes are intentional and covered by smoke tests.

General Code Hygiene

• Comments were added only where they explain non-obvious behavior.
• Modified and added files end with a single trailing newline.
• git diff --check passed.
• Comments and error messages are in English.
• Comments and error messages are complete sentences where applicable.

C++ Specific

• Code follows the repository's Google C++ style conventions to the extent verified manually.
• clang-format version 21.1.0 was run remotely and the final dry-run passed.
• N/A - clang-tidy was not run in this session; this PR does not add kernel code or template-heavy operator implementation code.
• Operator parameter order is preserved from the generated base operator signatures.
• No explicit exceptions are thrown by the added code.
• N/A - no kernel files or launchers were added.
• N/A - no new operator implementation classes were added.
• No raw new/delete was added by this PR.

Python Specific

• Python changes are limited to tests and generator code.
• ruff format and ruff check were run remotely and passed.
• Python smoke/full tests passed remotely.

Testing

• Relevant remote NVIDIA build and pytest results are recorded above.
• All supported remote platforms affected by this PR were tested and recorded in the table above.
• New public C++ API has a smoke test under tests/test_cpp_api.py.
• N/A - no new operator numerical tests were added.
• N/A - no bug-fix regression test is required.

Build, CI, and Tooling

• N/A - full pip install .[dev] was not used for the final full-platform runs; each container installed the required build/test packages and then used pip install . --no-build-isolation --no-deps.
• CMake configure/build/install paths were verified remotely.
• GENERATE_PYTHON_BINDINGS=ON pybind target still builds remotely.
• No new runtime dependency was added.

Documentation

• N/A - README-level public API documentation is deferred; this PR exposes the API through installed headers and smoke tests.
• Public API surface is exposed through installed headers and covered by smoke tests.
• N/A - no user-visible breaking change is introduced.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers were committed.
• No third-party code was added.
• No unsafe pointer arithmetic, uninitialized reads, or missing bounds checks were introduced.

[voltjia]

请列出几组通过这样暴露出的 API 的函数签名。

[voltjia]

Addressed the review concern around _musa_smoke_passes.

Root cause: the earlier Moore run was not pinned to a working card, so torch.musa.synchronize() hung and the added smoke guard filtered out musa, reducing coverage to CPU-only scale. With physical card 6 selected via MUSA_VISIBLE_DEVICES=6, torch.musa.is_available() reports true and the normal device discovery path works. I removed _musa_smoke_passes in 4f451af0.

I also reran unqualified pytest on all six platforms pinned to card 6 and updated the PR description with the full results. Moore is now 8471 passed, 7900 skipped; Ascend is 7405 passed, 8906 skipped with ('cpu', 'npu') discovered before the run.

[voltjia]

请 @crapromer 或 @wooway777 初审，@Ziminli 终审。

[wooway777]

回头把ci过一下吧，ci服务器g了。老黄说没问题反正