Last and best models are linked together after training

For some reason, last and best models are intermixed during checkpoint saving.

I wonder if this is not a problem with our logging configuration for lightning.

assigned to @andre.anjos

Check here: https://gitlab.idiap.ch/biosignal/software/mednet/-/blob/main/src/mednet/engine/trainer.py?ref_type=heads#L191

Version 2.2.0 of lightning is available - we should may be upgrade before...

mentioned in commit a3a75a36

mentioned in merge request !20 (merged)

mentioned in commit 37e0eeb1

closed with commit a3a75a36