Create train, evaluate and predict commands

Currently, the evaluation process is done directly after training, and there is no easy way (without commenting the training part) to run only the evaluation step.

We need to split the train_and_test function into two distinct commands: train and evaluate. We could also update the predict command so that it shares some code with the training and evaluation steps (instead of having a separate predict function in [prediction.py/DAN]https://gitlab.teklia.com/atr/dan/-/blob/1956acc9bbe79f6ce1b8ffba69fd3fe16cb49571/dan/predict/prediction.py#L92).