Add "final_extra_valid_opt_filepath" to train_model.py

[moyapchen]

I've been annoyed at having to kick off separate eval loops at the ends of trains, so here we are.

Also added a way to configure saving outputs to json (cause it's something that I've wanted for a while and I didn't want to add another return value in case scripts elsewhere were expecting only v_report and t_report).

Not super attached to names or to the default values I've got set. I tried using 'Optional[str]' for the cmdline arg, but the test was not happy with that...

Testing steps
Wrote a new test that exercises all of the new functionality.

[stephenroller]

Just want to finish discussion

[stephenroller]

Aren't the final evals stored in the trainstats file already? If not, my suggestion would be we include them in the trainstats.

[stephenroller]

And not even make it an option. Just do it.

[klshuster]

agree with all of @stephenroller's suggestions

[stephenroller]

I think we'd prefer keeping it with the rest of the .trainstats?

[moyapchen]

The other valid/reports aren't saved to .trainstats right now.

[stephenroller]

Every validation run, except the final one, is stored in trainstats.

ParlAI/parlai/scripts/train_model.py

Lines 474 to 481 in 9b5f8c6

	json.dump(
	{
	'parleys': self.parleys,
	'train_time': self.train_time.time(),
	'train_steps': self._train_steps,
	'total_epochs': self._total_epochs,
	'train_reports': self.train_reports,
	'valid_reports': self.valid_reports,

I'm humbly requesting you expand it to also include the final valid/test reports. (They can just be None during the intermediate stages)

[moyapchen]

Alright, wording wasn't quite clear there prior. Anyway, done.

[stephenroller]

Seems like you have a bug for when model file isn't set (and we should be throwing trainstats into the void). Otherwise lgtm, thanks for making the changes.