Update to latest structure and fix a variety of issues

This merge cleans up a lot of duplicated code across the stages. In particular there is now a universal "Mapping function" in mapping_utils.py, which takes any type of tensor (node-like, edge-like or track-like) and produces any other. It is quite robust to various input shapes.

Additionally, the problem of (un)directedness in graphs is handled for now by removing and duplication of edges. This happens now only inside a GNN module, and the output is to sum the edge scores. Thus graphs are left unaffected in input and output.

All stages now support feature scaling and unscaling.