disable/enable early stopping

Currently, only the best model on val data is stored and finally considered as the best model. Maybe it would be better to set during experiment setup, if the user wants to use early stopping or not.

• introduce parameter early_stopping_epochs: enable early stopping and set number of patience epochs
• introduce parameter restore_best_model_weights: use either the best model according to va loss (if True) or the model of the last train epoch (if False).
• be able to use the model status from last train step ✔

This relates to #367 (closed), as this should either mark the early stopping or the last step.