Add optimizer configuration to eval

[ohmeow]

Changes

Instead of hardcoded optimizers using in task specific train/eval runs, run_evals_from_checkpoints.py will now create a default optimizer configuration as well as default task specific overrides to be used.

Discussions

Discussed with @warner-benjamin so as to bring the finetuning configs more in line with the pretraining

Tests
No tests added

• Is the new feature tested? (Not always necessary for all changes -- just adding to the checklist to keep track)
• Have you ran all the tests?
• Do the tests all pass?
• If not, have you included an explanation of which tests this PR breaks and/or why (below this checklisT)

[ohmeow]

Ok I think this is good to go ...

[warner-benjamin]

There's one error which needs to be fixed, and am curious to hear if you have any thoughts on potentially avoiding the second set of hardcoded variables,

Otherwise looks good.

[warner-benjamin]

I'm not sure that adding a second set of hardcoded defaults is the best idea, but I'm not sure there's an easy way to support a dynamic input in typer,

[ohmeow]

One option is to remove the hardcoded optimizer defaults from the ClassificationJob classes. I think we want this driven in the config for two reasons:

1. It's clear to the user how the optimizer is setup without them having to go look in the the classes for each task.
2. It will make logging these things in wandb a bit easier (not 100% on this)