Upgrade to PyTorch

[PTRRupprecht]

CASCADE was built using the deep learning framework Tensorflow. To avoid dependencies on Tensorflow, it would be convenient to have a reimplementation of CASCADE in the probably most broadly used framework, PyTorch.

[PTRRupprecht]

For such an update, it was suggested by Mranmay Shetty to make another practically relevant change:

Right now model choice is mostly by name. It’d be great to add small per-model yaml metadata (indicator, fps, cell type, brain region, training dataset IDs) plus a simple auto-selector that recommends the best pretrained model given the user’s recording metadata (indicator, fps, cell type).