Draft: Try out new training for long electrons

I'll try out a new training to deal with long electrons, with the new version of the repository.

Here are the features that are expected to be implemented for the sake of this new training

• Be able to have more than 1 parquet file as input.
• Get the best result that one can from the embedding
• Better comparisons between train-val-test for embedding and GNN steps
• Use weights in the GNN that are inversely proportional to the number of clusters (1 / plane).
• Implementation of a new GNN after the default GNN, where nodes are doublets and edges correspond to triplets (see #5 (closed))
• Set all the edges towards increasing z instead of using a bidirectional graph (see #6 (closed))