Return training and val metrics in NoOpTrainer

[bryant1410]

Wouldn't it make sense for NoOpTrainer to return training and validation (and "best_validation") metrics (but not loss)?

For example, if I train several models and consume the TrainerBase API, I'd expect to get the training and validation metrics returned by the train method (e.g., accuracy, but not "loss" because it could possibly not make sense), seamlessly, regardless it's a Trainer or a NoOpTrainer implementation.

An alternative would be to always call evaluate on every set I want it to. But I think it's more practical if I could get the metrics of a no-op trainer right away, without calling evaluate.

Thoughts? I'm willing to send a PR if you folks think it's a good idea.

[brendan-ai2]

Thanks for the offer, @bryant1410! I've mostly been using the NoOpTrainer and then evaluating separately, but we think what you've outlined makes sense. We'd be happy to review your PR. One initial request, could you hide that feature behind a flag for when that feature is not required? Thanks!