Early stopping: save model only if validation loss or perplexity improves

[danielwatson6]

This would save time from re-training models until a specific number of steps (that must be computed manually since the number of training steps is determined by the number of epochs)