Refactoring changes in the training routine

[wiederm]

Description

This PR introduces a few improvements/changes to the training routine and its tests:

• improve logging messages
• improve fixture in conftest to obtain a single batch from different datasets and with different batchsize without hardcoding any of the values
• provide training toml file that sets parameters for force training
• The Error classes will return now the error per molecule instead of the mean of the error, this makes it easier to log and use torchmetrics
• tracking loss now also with torchmetrics instead of custom torch.nn.Module
• allowing to cache the processed dataset. In that case we assume that another process has already processed the dataset (since the .pt file is present) and, if cache regeneration is not requested, we will use this file and skiptt the prepare_dataset operation.
• lock the prepare_dataset method: only a single instance should execute this method per dataset.

This PR also includes bugfixes:

• fix a bug in the force loss calculation
• only retrain/create graph of force when in training mode (not in evaluation or test mode)

Status

• Ready to go

[chrisiacovella]

The CI was failing due to running out of memory (CI runner is capped at 16 gb). See #246 . The code will skip training sake with forces.