Summary

The stable PyTorch builds at pytorch,org have included builds for ROCm 6.4 since PyTorch 2.8. It would be great to add this to the --torch-backend logic, similar to what was done for cu130 in #16321.

I'm not sufficiently expert to generate a PR myself, which is why I'm filing this as an issue.

Example

Makes the uv pip install torch --torch-backend=auto more useful by adding awareness of recent stable builds for AMD GPUs