Load best model + evaluate best model on devset

[OrianeN]

I've noticed that the model loaded at the end of training differs whether the EarlyStoppingException was raised or not:

• if EarlyStoppingException: load best model
• otherwise, keep on with the last model

I'd like to suggest always loading the best model at the end of the training phase, as well as performing the final evaluation successively on the devset and on the testset (with the best model).

Logs of 2 runs to show you the changes in practice:

• without loading the best model: papie_run_noloadbest.log (POS loss of the last model is higher than the one-but-last one)
• with loading the best model: papie_run_loadbest.log

Note: I've never tried training without a dev set, so I'm not sure about the comment at the end of the Trainer.train_epochs() function:
https://github.com/OrianeN/PaPie/blob/d23f4a5e95638e74dad640cf7f1d44b896c2fb0e/pie/trainer.py#L429-L430

[OrianeN]

As discussed, I made the training script save both the best and the last model, as well as print the index of the best epoch (starting from 1).

Here are the updated example logs: papie_run_loadbest_saveboth.log