Not saving best model

[andreped]

The current checkpoint callback saves a model at a given step periodically. However, it is not necessarily the best model.

We should use the EvalCallback: https://stable-baselines3.readthedocs.io/en/master/guide/callbacks.html#evalcallback

[andreped]

This was added here, but it fails due to issue #3.

As training occationally crash and we are unable to properly update the environment, forcing us to completely restart training, the EvalCallback will have the wrong behaviour.