Nova builds CUDA wheels on GPU runner

[huydhn]

A timeout error when building FBGEMM CUDA wheel https://github.com/pytorch/FBGEMM/actions/runs/10772363019/job/29869844126 uncovers the fact that Nova builds CUDA wheels on GPU runners https://github.com/pytorch/test-infra/blob/main/tools/scripts/generate_binary_build_matrix.py#L120. This isn't the most efficient way to use these runners, but this was done this way because domains builds were quick and didn't need a separate test job. The assumption doesn't hold for FBGEMM.

Solution: Check if it's worth the effort to refactor Nova build job into build and validation parts, only the latter requires GPU while we can build on a bigger CPU runner.

(We also need this when we want to use Nova in PyTorch in the future, I'm creating this to track the issue)

cc @atalman @malfet @spcyppt