Provide models with pre-computed neighbor lists during training

[Luthaf]

This is split from #174 (comment). While discussion the interface for model training I though it might make sense to provide the models with pre-computed NL when training.

We already pre-compute these in inference mode, so doing the same for training mode makes sense to me. The interface could looks something like

class ModelInterface:
    ...

    def requested_neighbor_lists(self) -> [NeighborListOptions]:
        ...  # often the same set of NL as the exported model

# ==================================================================================== #

# training loop

if "continue_from_checkpoint":
    model = __MODEL_CLASS__.from_checkpoint("path")
    model = model.restart(dataset_info)
else:
    model = __MODEL_CLASS__(hypers["architecture"], dataset_info)

# NEW CODE HERE
for request in model.requested_neighbor_lists():
    for system in [train_datasets, validation_datasets]:
        neighbors = compute_nl(request, system)
        system.add_nl(request, neighbors)
# END OF NEW CODE

trainer = __TRAINER_CLASS__(hypers["train"])

trainer.train(
    model=model,
    devices=[...],
    train_datasets=[...],
    validation_datasets=[...],
    checkpoints_dir="path",
)