WIP: [Torch] Renamed Torch to TorchCUDA, and added a CPU-only Torch

[stemann]

Currently a WIP. Aims at resolving FluxML/Torch.jl#20.

Similar approach as for ONNXRuntime - separate binaries for CPU-only and CUDA etc. (CPU: #4369, CUDA: #4386).

Relates to: #1529

[vchuravy]

Does this need to be a different package? Or can we turn this into a variant?

[stemann]

Does this need to be a different package? Or can we turn this into a variant?

If eventually both CPU, CUDA and ROCm builds are available, wouldn't the different sets of dependencies necessitate different packages?

[vchuravy]

You still have the same issue that if Torch.jl includes Torch_jll, TorchCUDA_jll,... all dependencies must be fulfilled. Better to use a platform tag and platform specific dependencies.

[stemann]

You still have the same issue that if Torch.jl includes Torch_jll, TorchCUDA_jll,... all dependencies must be fulfilled. Better to use a platform tag and platform specific dependencies.

Right, I see. Is there some discussion/registry of defined/reserved platform tags? I.e. that "cuda" is reserved for CUDA and so on for e.g. "amd_rocm", "intel_oneapi", "apple_coreml", ...

[stemann]

Superseded by #4554