dev2main1

• write torch.cat((tensor1, tensor2)) instead or torch.cat([tensor1, tensor2]), because it works better with typehints
• Be able to run build_embedding on CPU, which is faster because of the kNN
• Rename building parameter in metric_learning section to processing
• Create GNNLazyDataset in order to load the events during GNN training and apply a processing beforehand. Here, I apply a processing for bidirectional graph, which we will be removed later because I don't care about bidirectional anymore
• Define on_step hyperparamameter to log the loss each step instead of each epoch, just for debugging. Probably useless.
• Define DataFrameLoader during preprocessing. Even though the 700,000 events scattered across 350 folders, this loop allows to loop over the events, kind of as if they were belonging to a single file.
• Define compute_n_unique_planes custom processing
• Define step 7 to compare ETX4VELO with Allen
• Be able to load events in a lazy way, implemented in ModelBase.