Run x86 CUDA wheels on HF Jobs

[andimarafioti]

Summary

• Route both Linux x86_64 CUDA 12.8 and CUDA 13.0 wheel builds to hf-jobs-cpu-performance.
• Keep Linux aarch64 CUDA builds on ubuntu-24.04-arm, because the HF Jobs runner labels select hardware flavors, not ARM64 Actions hosts.
• Increase x86_64 CUDA native build parallelism from 2 to 16 jobs on the HF CPU performance runner.

Notes

This uses the Hugging Face Jobs Actions bridge, where GitHub queues jobs with runs-on: hf-jobs-*, an ephemeral HF Job runner picks them up, and GitHub dispatches the queued workflow job to that runner.

Relevant docs:

• https://github.com/huggingface/jobs-actions
• https://huggingface.co/blog/github-ci-hf-jobs
• https://huggingface.co/docs/huggingface_hub/en/guides/jobs

Validation

• Parsed .github/workflows/wheels.yml with PyYAML and confirmed both x86_64 CUDA matrix entries use hf-jobs-cpu-performance with 16 build jobs.
• git diff --cached --check

I did not run a full wheel build locally; this PR is intended to exercise the new runner routing in GitHub Actions.