python3Packages.torchWithRocm: 2.0.0 unnecessarily requires unfree dependencies

[SomeoneSerge]

Issue description

Once we update python3Packages.torch to 2.X.X, its torchWithRocm variation will depend on cudaPacakges.cuda_nvcc (through python3Packages.openai-triton), which is unfree. This means hydra won't be building and caching torchWithRocm.

Upstream enforces that there be a copy of bin/ptxas (from cuda_nvcc):

• https://github.com/pytorch/pytorch/blob/558e5a240e0c2a5e9a0d5350fe6c25b029efca6a/setup.py#L931
• https://github.com/openai/triton/blob/01a93185a1105cf4b3ab504ec385ff3585160932/python/setup.py#L211

However, they do not actually use it except for cuda support (obviosuly).
We should patch around this to re-enable caching, and we should work with upstream to make the dependency optional in the first place

CC @NixOS/rocm-maintainers

[Madouura]

Sadly, with the new ROCm updates, I've only made it worse.
We can't seem to reasonably patch hsa-amd-aqlprofile out of rocprofiler anymore.

[Madouura]

This was fixed in 6f3f5b6 and #263048.