[SYCL][Fusion] Kernel Fusion support for CUDA backend

[sommerlukas]

Extend kernel fusion for the CUDA backend.

In contrast to the existing SPIR-V based backends, the default binary format for the CUDA backend (PTX or CUBIN) is not suitable as input for the kernel fusion JIT compiler.

This PR therefore extends the driver to additionally embed LLVM IR in the fat binary if the user specifies the -fsycl-embed-ir during compilation, by taking the output of the sycl-post-link step for the CUDA backend.

The JIT compiler has been extended to handle LLVM IR as input format and PTX assembly as output format (including translation via the NVPTX backend). Target-specific parts of the fusion process have been refactored to TargetFusionInformation.

The connecting logic to the JIT compiler in the SYCL RT has been extended to produce valid PI device binaries for the CUDA backend/PI.

Heterogeneous ND ranges are not yet supported for the CUDA backend.

[sommerlukas]

/verify with intel/llvm-test-suite#1683

[sommerlukas]

@sergey-semenov: The changes to graph_builder.cpp and commands.hpp are necessary to avoid deletion of dependencies.

With fusion, the individual commands for each kernel are replaced by a single command for the fused kernel, and the original commands are deleted without execution. Without this modification, the destructor of the command would call cleanDepEventsThroughOneLevel, which would not only delete the dependency edges of the original command, but also its dependencies.

This would yield an incomplete dependency graph. Instead, the dependencies of the original command are deleted before removal, but only in the case a command is removed without execution during fusion.

[steffenlarsen]

Both the runtime and design doc changes LGTM. Only a couple minor nits.

[bader]

@intel/dpcpp-clang-driver-reviewers, could you review driver's changes, please?

[mdtoguchi]

OK for driver

[sommerlukas]

@intel/llvm-gatekeepers This is now approved, could someone please merge this?

[aelovikov-intel]

I'm seeing

Unexpectedly Passed Tests (1):
  SYCL :: KernelFusion/device_info_descriptor.cpp

in CUDA pre-commit CI tasks on unrelated PRs. @sommerlukas would you take a look at it, please?

[sommerlukas]

I'm seeing

Unexpectedly Passed Tests (1):
  SYCL :: KernelFusion/device_info_descriptor.cpp

in CUDA pre-commit CI tasks on unrelated PRs. @sommerlukas would you take a look at it, please?

@aelovikov-intel Is pre-commit CI running with the latest version of the e2e tests, specifically the device_info_descriptor test? Prior to this PR, the test was marked XFAIL for cuda, but that was removed in this PR (only XFAIL on hip now).

If the version of the e2e tests in pre-commit CI still expects XFAIL for cuda, but the test now passed, that would explain the unexpected pass on CUDA.