Improved training tracking: compute metrics, save predictions

[jdeschamps]

Problem

Currently, only the loss on validation images is reported during training. We should allow the possibility to compute metrics (scale invariant PSNR, MicroSSIM) if applicable, and save predictions on a validation image.

Ideally, this would be compatible with the three supported logger:

• CSVLogger: metrics should be reported.
• WandB: metrics and prediction should be saved.
• TensorBoard: same as WandB.

Finally, predictions could be saved to a user-defined folder.

Potential solution

Firstly, we now have the capacity to predict during training, although the solution is pretty hacky: #266.

Secondly, PyTorch Lightning should have all the engineering required to save the metrics and the predictions in the relevant loggers:

• Logging from a LightningModule
• Log media in WandB
• TensorBoard

Linked issues

#299: better convenience functions for N2V