pytorch CUDA support within key4hep environment

[dudarboh]

The current key4hep builds pytorch without CUDA support.
That means all HEP ML studies where one would need to train on GPUs are performed in a separate custom user environment.
I think it would greatly help the reproducibility and integrability of all ML studies if training on GPUs could be possible within the same key4hep environment where one runs all the other non-ML related things, e.g. simulation/reconstruction/analysis/etc.

As discussed privately with @tmadlener, including CUDA support for all O(10) CUDA architectures would increase the stack size by 4-5 GB (~30% of its current size) and add a lot of maintenance burden.
So it may only make sense if many people are interested.