How should we make use of `__cuda_arch` / the `cuda-arch` package

### Comment:

At prefix we are quite interested in the pytorch package.

We have been thinking on how to support the new `__cuda_arch` capabilities, especially for large packages like `pytorch`. Targeting more specific CUDA architectures should shrink the package size.

One thought was that maybe the pytorch feedstock could create multiple architecture specific variants, and one "fat" variant for backwards compatibility when the Conda/Mamba/Pixi tool doesn't know about `__cuda_arch` yet.

Are there any other thoughts in the community about this? @carterbox can you sketch out what you concretely envision for packages like `pytorch`?

Ref: https://github.com/conda-incubator/nvidia-virtual-packages/