Training progress bars

[Dschogo]

I fiddled a bit with training, could look like this (including multi GPU - theoretical couldn't test) - Updates ~ 1 a sec

This wont be included in this PR - if its desired to add it also in training, then I'll try to polish it a bit more because the logger breaks it here and there:

The progress redraws itself again, and it affects nothing in the back-end. Just a minor visible annoyance.

Since the training data was previously printed through the logger, it will no longer be logged.

[Dschogo]

Seems to work fine (trained 5 models from ~ 15min audio for 100 epochs - no issues.
It should work for mutliple GPUS, maybe its a bit more buggy (only terminal progress bar, the training will work fine)

Also changed the "save only last epoch" method to actually save the new epoch, after a successful save it deletes the previous highest iteration counter Regex: G_*.pth / Regex: D_*.pth