Overhaul of Adeft validation

[steppi]

This PR makes an overhaul of validation in adeft it

• Uses a proper nested cross validation scheme to isolate model selection from validation.
• Uses pooled F1 for model selection. This means of computing F1 across cross validation folds is recommended in the classic Apples to Apples paper by Forman and Scholz (https://dl.acm.org/doi/10.1145/1882471.1882479).
• For validation results, reports classwise sensitivity and specificity, and the full confusion matrix, rather than per class F1 and global F1. F1 wasn't actually a very good metric to report because the class balance in the training data doesn't reflect the actual class balance. Sensitivity and specificity are independent of the unknown class balance, so still readily interpretable. Future work will go into better understanding the class balance in unseen data.
• Corresponding updates had to be made for model serialization, info reporting in the AdeftDisambiguator and tests.

This PR is a work in progress. The tests are not going to pass until I populate the 1.0.0-dev folder on S3 with testing resources. I'm just posting this for the sake of visibility. Note also that this is a PR into a v1.0.0-dev branch. I may back down from committing to a stable 1.0.0 release, but it seems like a worthy goal to shoot for.

[steppi]

Shelving this for now in favor of a more incremental approach to more quickly get us where we need to be for the OPAQUE paper.