feat(hip-directed): add hip_directed catch2 e2e tests

[y1k2chaitanya]

Summary

Ports the hip_directed catch2 directed-unit-test suite into the public rocm-tests framework, mirroring the qa-forge rocm-func-core-hip-catch2-amd / Rocmtest hip_directed run.

• New tests/e2e/hip_directed/ area (category profile: hw.gpu, layer.runtime, ci.nightly, e2e.stack, os.linux).
• Clones public ROCm/rocm-systems at develop (overridable via ROCM_TEST_ROCM_SYSTEMS_REF), builds projects/hip-tests/catch, and runs a curated set of directed Unit_hip* tests via ctest -R (13 parametrized rows).
• Adds tests/e2e/hip_directed/ to the e2e-tests.yml pytest dir list so nightly CI collects it.

Directed tests (13)

hipSetValidDevices, hipGetDriverEntryPoint, hipStreamGetId, hipStreamSetAttribute, hipStreamGetAttribute, hipMemPrefetchAsync_v2, hipMemAdvise_v2, hipModuleGetFunctionCount, hipModuleLoadFatBinary, hipMemcpy3DBatchAsync, hipMemcpy3DPeer, hipMemcpyBatchAsync, hipMemsetD2D.

Build design

• Builds only the catch2 executables that hold the directed tests — DeviceTest, StreamTest, MemoryTest1, ModuleTest — instead of the whole build_tests meta-target. This skips modules like coopGrpTest (which need bleeding-edge HIP headers not in every ROCm install) and keeps the build to seconds instead of the full ~4095-test build.
• Limits GPU visibility during the build (ROCR_VISIBLE_DEVICES=0): catch2's CMake auto-detects the offload arch via rocm_agent_enumerator, which returns one entry per visible GPU — on multi-GPU hosts that duplicates --offload-arch and trips clang-offload-bundler: Duplicate targets. A single visible device yields one arch.
• Per-test rows: a build/run failure in one module doesn't block the others.

Test plan

Validated on MI325X (gfx942):

• pytest tests/e2e/hip_directed/ --rock-dir <rocm> --gpu-arch gfx942 -q
• 13/13 passed (~59s incl. the scoped build)
• --collect-only shows 13 tests
• ruff + black

[y1k2chaitanya]

Local validation — quanta MI325 (gfx942), ROCm 7.14.0-dev0+22b25e9

Pinned rocm-systems to the exact commit the installed ROCm was built from — read from <rock_dir>/share/therock/therock_manifest.json (pin_sha 10de68988e384c36d5a076a14d7c14f8e49c6a79), with ROCM_TEST_ROCM_SYSTEMS_REF override and develop fallback. This keeps the hip-tests catch2 source in sync with the installed HIP headers (fixes the version skew we hit against develop).

tests/e2e/hip_directed  →  12 passed, 0 failed, 0 skipped   (29.7s, build cached)

All 12 directed tests green on the manifest-pinned commit:
hipSetValidDevices, hipGetDriverEntryPoint, hipStreamGetId, hipStreamSetAttribute, hipStreamGetAttribute, hipMemPrefetchAsync_v2, hipMemAdvise_v2, hipModuleGetFunctionCount, hipMemcpy3DBatchAsync, hipMemcpy3DPeer, hipMemcpyBatchAsync, hipMemsetD2D.

Dropped hipModuleLoadFatBinary — its positive test loads build-generated .code fatbins produced by add_custom_target(... ALL ...) steps that are not dependencies of the scoped ModuleTest executable build (the negative test passes; not a product bug).

Pushed as 4c13820.

[y1k2chaitanya]

CI validation — e2e-tests.yml on linux-gfx94x-MI325 (gfx942)

Ran the suite through the e2e CI workflow (isolated via -k hip_directed):

collected 292 items / 280 deselected / 12 selected
12 passed, 0 failed, 0 skipped in 216.09s

Run: https://github.com/ROCm/rocm-tests/actions/runs/29980833297

Fixes made portable to the OSSCI container (pushed in 3daad92):

• compiler_mode=cxx_hip — sets CMAKE_HIP_COMPILER to the ROCm clang++ (a bare system clang++ cannot find the ROCm device library).
• Pin CMAKE_HIP_ARCHITECTURES — stops CMake's HIP ABI check from duplicating --offload-arch per GPU (rocm_agent_enumerator reads sysfs, ignoring ROCR_VISIBLE_DEVICES) on multi-GPU runners.
• Bundled sysdeps numa — find_library(numa)/find_path(numa.h) resolve from <rock_dir>/lib/rocm_sysdeps and MemoryTest1 loads librocm_sysdeps_numa.so.1 via LD_LIBRARY_PATH; no host apt packages required.