Build CUDA-enabled Python wheels on Linux and Windows

[zfergus]

Description

Builds the ipctk wheels with CUDA on Linux x86_64 and Windows, and adds a runtime check for whether the GPU code can run. macOS wheels stay CPU-only.

• Toolkit: CUDA 12.8. CUDA 13.x would require driver 580 or newer; 12.x needs 525 or newer.
  • Linux: cuda-nvcc, cuda-cudart-devel, and cuda-cccl from NVIDIA's rhel8 repo, installed inside the manylinux_2_28 container (GCC 14).
  • Windows: Jimver/cuda-toolkit, then msvc-dev-cmd, and the build uses Ninja instead of the Visual Studio generator, as cuda.yml does.
• Architectures: 75-real;80-real;86-real;90-real;100-real;120-real;120-virtual. Turing through Blackwell get SASS, and newer GPUs can JIT from the sm_120 PTX.
• Runtime: static cudart (the CMake default). The wheel needs only the NVIDIA driver, and auditwheel has no libcudart to bundle.

Impacts:

• pip install ipctk gets ipctk.cuda.LBVH on Linux and Windows.
• On a machine without a GPU or driver, the wheel still imports. Using a GPU class raises RuntimeError.
• Windows wheels now build on windows-2022, because windows-latest ships MSVC 19.51 (VS 2026), which CUDA 12.8 rejects.

API changes

• New ipc::cuda::is_available() (ipc/utils/cuda/is_available.hpp) and ipctk.cuda.is_available().
  • Returns true when cudaGetDeviceCount succeeds with at least one device.
  • Always defined; returns false in non-CUDA builds.

CI

• pypi.yml now runs on pull requests that touch the wheel build. The upload_pypi job is skipped for them.
• Each Linux and Windows wheel runs python/tests/check_cuda_wheel.py on the GPU-less runner. It asserts the wheel is a CUDA build, that is_available() is False, and that LBVH.build raises instead of crashing.
• A step lists each wheel's size in the job summary and fails if any wheel exceeds PyPI's 100 MB limit.

Type of change

• New feature (non-breaking change which adds functionality)

How Has This Been Tested?

• [utils][cuda] Catch2 test: in CUDA builds, is_available() predicts whether cuda::LBVH builds and detects or throws. In CPU builds, it is false. cuda.yml runs the GPU path on its simulated T4.
• python/tests/test_cuda.py: the same behavior from Python. python.yml runs it on CPU builds.
• Locally (macOS, Apple Clang, Release, CPU-only): both tests pass, and check_cuda_wheel.py correctly rejects the CPU build.
• pypi.yml CUDA wheels build and pass the no-GPU check (this PR's CI).

Checklist

• My code follows the clang-format style guidelines of this project
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.92%. Comparing base (4e8173d) to head (3fba971).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #274   +/-   ##
=======================================
  Coverage   96.92%   96.92%           
=======================================
  Files         194      194           
  Lines       17349    17349           
  Branches      943      943           
=======================================
+ Hits        16815    16816    +1     
+ Misses        534      533    -1     

Flag	Coverage Δ
unittests	96.92% <ø> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.