[trainer] Add a loss recorder option

[frgfm]

Once a model has been trained, it would be nice to analyze its main failures. For that, one could:

• go through the training set computing the loss (unreduced on the batch dimension)
• keep the N samples with the highest loss
• plot the image, the loss, the prediction and the target

A later option would be to display the CAM to better understand the model behaviour.