Remove outdated evaluate code and just use trainer.validate.

[bw4sz]

This PR

• Removes __evaluate__ and calculate_empty_frame accuracy
• uses trainer.evaluate in main.evaluate()
• Keeps the output structure that user is used to
• stashes the .results object in the trainer.

closes #1316

and simplifies the code.

Thanks to @jveitchmichaelis for his idea here.

[bw4sz]

Okay, @jveitchmichaelis I think this ready to review. I removed predictions arg from the evaluate CLI, it was a weird edge pattern that didn't match the rest of the workflow.

[bw4sz]

I reverted the .toml, it was specific to my laptop, let's make that a separate PR.

[jveitchmichaelis]

Looks good aside from that docstring + needs squash.

I do like the idea of being able to run eval on saved predictions, but I agree there's probably a better place to do it (and it feels awkward here if we standardize on using the trainer).

[bw4sz]

rebased.

[jveitchmichaelis]

LGTM!