[Build] Only build 9.0a for scaled_mm and sparse kernels

[LucasWilkinson]

Only build 9.0a for scaled_mm and sparse CUTLASS kernels to save on build time and binary size (instead of 9.0 AND 9.0a). Since these kernels use 9.0a specific features and theres no pure 9.0 systems that we are aware of, only the 9.0a versions are getting used.

Also clean-up the cmake utility fixing a bug where cuda_archs_loose_intersection was ignoring the TGT_CUDA_ARCHS parameter and just using CUDA_ARCHS directly

Wheel size:
this PR - 191.42 MB
main - 191.46 MB

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀