Make it easy to save training curves locally

[louisfh]

A user pointed out that though wandb integration is nice, they'd like to be able to output training and validation metrics locally, instead of being forced to use wandb.

Maybe we could expose an argument to train that would enable the user to save training and validation performance. Or maybe we just point people to wandb.

[sammlapp]

I agree. I like the flexibility Lightning gives, where you pass a Logger which can be a local file logger, Tensorboard, WandB etc. We might be able to mirror/leverage Lightning logger flexibility. However a lighter weight solution that I'm already planning for 0.13.0 is to provide more local training outputs, with inspiration from YOLO. The branch @visualization creates html elements with grids of embedded audio-spectrogram. I envision creating an html page (saved locally) with training results like loss graph, metrics vs step, and preprocessed samples.