CI: CUDA version mismatch in 13.0.2 local test jobs #1691

Summary

The 13.0.2, local CI test jobs install nvidia-cuda-runtime-13.1.80 via pip (pulled in by cuda-toolkit[cudart]==13.*) alongside a local CTK 13.0.2. This creates a version conflict: the local nvjitlink is 13.0.88 but libcudadevrt.a from the pip wheel is 13.1.80.

How it was discovered

PR #1690 updates test_device_launch.py to use cuda.pathfinder.find_static_lib("cudadevrt") instead of a manual CUDA_HOME-only lookup. The pathfinder searches site-packages first (by design, matching all other pathfinder finders), so it returns the 13.1 libcudadevrt.a from the pip wheel. When nvjitlink 13.0 tries to link it:

cuda.bindings.nvjitlink.nvJitLinkError: ERROR_INVALID_INPUT (3)
Linker error log: ERROR 4: bad input:/opt/hostedtoolcache/Python/3.13.12/x64/lib/python3.13/site-packages/nvidia/cu13/lib/libcudadevrt.a

The old test code avoided this by only looking at CUDA_HOME, which pointed to the local 13.0.2 CTK.

Affected jobs

• py3.13, 13.0.2, local, h100
• py3.13, 13.0.2, local, rtxpro6000
• py3.10, 13.0.2, local, rtxpro6000 (TCC) (Windows)

Why other jobs pass

• 13.1.1, local jobs: the local CTK and pip wheel are both 13.1.x, so no mismatch.
• 13.0.2, wheels jobs: the device launch test is skipped due to architecture (sm_89 < sm_90).
• 12.9.1, local jobs: no cu13 pip wheel conflict.