Add an easy way to compute loss during validation

[KevinMusgrave]

It can be useful to compute the average loss over the training set or validation set, and use that (rather than some accuracy metric) to select the best checkpoint. See #122 (comment) for a discussion.