pathfinder: add sm_arch bitcode lookup

[aryanputta]

Summary

• add optional sm_arch lookup to locate_bitcode_lib() and find_bitcode_lib()
• normalize accepted SM forms like sm90, sm90a, sm_90, and 90 to the package filename suffix
• enable arch-specific lookup for nvshmem_device as libnvshmem_device_{sm_arch}.bc
• cover normalization, search order, invalid input, and unsupported-library behavior in pathfinder tests

Fixes #2200

Validation

• python3 -m py_compile cuda/pathfinder/_static_libs/find_bitcode_lib.py tests/test_find_bitcode_lib.py
• git diff --check
• awk 'length($0) > 120 { print FILENAME ":" FNR ":" length($0) }' cuda_pathfinder/cuda/pathfinder/_static_libs/find_bitcode_lib.py cuda_pathfinder/tests/test_find_bitcode_lib.py
• pre-commit run ruff-check --files cuda_pathfinder/cuda/pathfinder/_static_libs/find_bitcode_lib.py cuda_pathfinder/tests/test_find_bitcode_lib.py
• pre-commit run ruff-format --files cuda_pathfinder/cuda/pathfinder/_static_libs/find_bitcode_lib.py cuda_pathfinder/tests/test_find_bitcode_lib.py
• graphify update .

python3 -m pytest tests/test_find_bitcode_lib.py is blocked in this local macOS environment during collection before these tests run. Importing cuda.pathfinder selects the Linux dynamic-loader module and fails on missing dlinfo from /usr/lib/libdl.dylib.

[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.

[rwgk]

There is already #2201. It's currently on hold, pending internal discussions. I'll close this PR.

BTW: I started reviewing your #2196 but was too busy to finish the review. One high-level observation: it's a good start, but fixing up the Gen-AI-isms (e.g. over-explaining things that are not relevant beyond the scope of the PR) directly is far easier these days than crafting review comments. I'm still figuring out how to deal with that best.