Use python logging in training.

This way, we get the training logs in the experiment_root too!
TODO: Maybe also do that in embed and eval?