Refactor Dataloader output

[vitusbenson]

Dataloaders currently output a dict with keys like "highresdynamic" or "mesodynamic" etc. This is intransparent and looses a lot of metadata.

Goal:

• Dataloader output should maintain metadata (what is which variable, which lat/lon/time step, which sample etc.)
• Be more flexible for different datasets
• Still use PyTorch Tensors, since ultimately we are restricted here by PyTorch Lightning

Risks:

• Backward compatibility will likely break by this