Release single-step evaluation framework and wrappers for several model types

[kmaziarz]

This (fairly large) PR releases the majority of the single-step evaluation code, including the eval script itself, inference wrappers for seven model types, and glue code to plug these models into search. The changes are split into semantically meaningful commits. Some notes for reviewing:

• Wherever the commit message starts with "Release", this means our internal code was copied verbatim (or almost verbatim), and thus at the very least it went through code review in the past.
• Some of the commits edit the configuration around CI and pre-commit. These changes were needed partly because the model wrappers contain dependencies which are not installed in our default environment, and thus one needs to tell mypy to ignore external libraries it doesn't know about.
• Some bits of internal code are not part of this release; the "Make the evaluation pipeline easy to extend" commit makes it easier to plug them in into syntheseus from outside.

A few notable things which are still missing, to be added in future PRs to avoid making this one even larger:

• Detailed setup instructions for each model type (e.g. environment definitions).
• Model checkpoints for each model type.
• More tests covering syntheseus/reaction_prediction (currently the coverage is quite low, partly because the wrappers themselves cannot be tested easily without setting up dedicated environments for each, but there are also places which are just not adequately covered by tests yet).

[AustinT]

I left a few small comments, but overall looks good to me. The tests all pass locally for me.

I was not able to run eval.py locally because I don't have the correct model files or example configs, so I can't independently confirm that it works. By extension, I could not test the "Make the evaluation pipeline easy to extend" commit. Did you test it locally @kmaziarz ? If so I am happy to merge it, and provide documentation / configs / model weights in a future PR.

[kmaziarz]

I was not able to run eval.py locally because I don't have the correct model files or example configs, so I can't independently confirm that it works. By extension, I could not test the "Make the evaluation pipeline easy to extend" commit. Did you test it locally @kmaziarz?

Yes, I re-ran the single-step evaluation for all seven model types, and confirmed that all resulting numbers match the table in the paper 🙂