Include cuda.core C++ .h files in packages

[fallintoplace]

Summary

• Include _cpp/*.h in the cuda-core source distribution manifest.
• Include _cpp/*.h in setuptools package data so wheel builds carry the C++ helper header.

Why

cuda/core/_tensor_map.pyx includes _cpp/tensor_map_cccl.h, but the _cpp packaging metadata only included .hpp headers. Builds from packaged artifacts could miss the required header.

Checks

• cd cuda_core && uv run --no-project --with build python -m build --sdist --outdir ../cuda-python-package-check
• tar -tzf ../cuda-python-package-check/cuda_core-*.tar.gz | rg "cuda/core/_cpp/tensor_map_cccl\.h$"
• cd cuda_core && uv run --no-project --with "setuptools>=80" --with "setuptools-scm[simple]>=8" --with "Cython>=3.2,<3.3" --with "cuda-pathfinder>=1.5" python setup.py build_py --build-lib ../cuda-python-build-py-check
• test -f ../cuda-python-build-py-check/cuda/core/_cpp/tensor_map_cccl.h
• git diff --check -- cuda_core/MANIFEST.in cuda_core/pyproject.toml

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[mdboom]

Makes sense to me. Why did our sdist testing not catch this? Is there something we should improve there?

[mdboom]

/ok to test 83ee9cf

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2236/
https://nvidia.github.io/cuda-python/pr-preview/pr-2236/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2236/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2236/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

/ok to test ed41580

[leofang]

Approving. Quick survey while checking whether this is a real gap (and to answer @mdboom's sdist question).

What's in cuda-core 1.0.1 wheel today (cp312, manylinux_x86_64; same shape across builds): 0 .pyx, 0 .cpp (stripped — implementations live in the .so); 67 .pxd, 15 .h/.hpp (kept — declaration surface for downstream cimport/#include). _cpp/resource_handles.hpp is shipped; _cpp/tensor_map_cccl.h is silently missing.

Why it hasn't failed: test-sdist-linux builds sdist → wheel-from-sdist, but setuptools-scm auto-bundles every git-tracked file into the sdist regardless of MANIFEST.in, so the wheel build never sees a missing header. And at runtime the compiled _tensor_map*.so already contains the symbol; _tensor_map.pxd has no cdef extern from, so no downstream cimport chain reaches this header either. Purely a packaging inconsistency, not a broken wheel.

De facto rule this PR upholds: for _cpp/ and _include/, ship declaration files (.pxd, .h, .hpp) and strip implementations (.pyx, .cpp). main had "cuda.core._cpp" = ["*.hpp"], inconsistent with the sibling "cuda.core._include" = ["*.h", "*.hpp"] and dropping .h files from _cpp/. This PR restores the pattern.

[leofang]

/ok to test fe08af4