Option to write predictions to disk

Option to write predictions to disk

[jeremyasapp]

Yeah I think a boolean on the Trainer and Evaluator as you've suggested offline makes sense to me. Will work on that after #83

Don't mean to harp on it but... I see this being significantly less useful without ids 😬 Because then you simply can't shuffle = False, nor drop_last = True, in BaseSampler, right?

Actually, meh, nvm.

Let's say that we get 2/3 of the usefulness from just writing to disk, and the remaining 1/3 with ids.

With just the former, we can still do all the global aggregations.

[cle-ros]

That's a great idea! However, while we're at it: would it be possible to make this more general, and more user-controllable? I.e., could we maybe have some general "results-writer" that can be triggered for some things automatically, but that could also be included in arbitrary models?

E.g., something like an automatic boolean called "log" that's a wrapper around each model s.t. the input-output pairs of the nn.module will be stored? But additionally having something like a results_writer object?

[jeremyasapp]

@cle-ros That s a sweet idea. I'll think about it after the next release!