Updating training script

[wiederm]

Description

This PR prepares modelforge with the infrastructure to effortless train the implemented neural network potentials with a single training script using pytorch-lightning.

Todos

• [x] Make the treatment of devices consistent
• [x] Fix bug in FirstComeFirstServeSplitter
• [x] update various type hints
• [x] make atom index mapping ~O(1) instead of O(N)
• [x] add useful options to lightning training

Status

• [x] Ready to go