Torch/TF2 version

MIDASpy is currently implemented using logic of TF1 and compatibility layers. As TF2 matures and more graph-based features become deprecated (see e.g. #21), we will need to plan for larger scale update of codebase.

We could try rebuild in TF2 natively or alternatively pivot to PyTorch implementation, which has a more "pythonic" feel.

