Vendor Clang resource headers for ast_canopy wheels

[isVoid]

Context

Numbast MLIR CI exposed that statically linking Clang into ast_canopy is not enough for a reliable PyPI install. CUDA parsing still needs Clang resource headers such as __clang_cuda_runtime_wrapper.h. If clang++ is missing at runtime, ast_canopy can fall back to a non-existent resource directory and fail before parsing user headers.

TODO

• Vendor the required Clang resource include directory from the LLVM/Clang version ast_canopy is built against.
• Make ast_canopy prefer the packaged resource directory when constructing -resource-dir=....
• Keep clang++ -print-resource-dir as an override/fallback path when explicitly requested or available.
• Include LLVM license and notice material in the sdist/wheel for the vendored headers.
• Preserve SPDX/license comments in vendored LLVM headers and document the source LLVM revision.
• Add packaging checks that verify __clang_cuda_runtime_wrapper.h and dependent resource headers are present in built wheels.
• Keep CUDA Toolkit headers out of the wheel unless separately reviewed; CUDA headers should continue to come from CUDA packages or CUDA_HOME.

Notes

LLVM/Clang resource headers are licensed under Apache-2.0 WITH LLVM-exception, which supports redistribution with retained notices. NVIDIA CUDA Toolkit headers have separate licensing and are not covered by this TODO.