Triplet GNN: `compute_loss`, `masked_edges` and new triplet building

This MR introduces

• Skip batches that go out of memory (for SciFi training)

• compute_loss argument in common_training_validation_step: only compute the loss when required. For instance, computing the loss in the test sample might require special columns, not available for the latter.

• masked_edges: a new hparam that allow to mask some edges

• A refactored triplet building from edges, that separates

  • building triplets from edges
  • computing whether a triplet is fake or genuine

  Allow to compute the truth only if necessary. This is also supposedly faster with less memory consumption.