Relearn evaluation and relevant results.

[Davidyz]

1. Script for automated relearning evaluation that automatically downloads models from hf repo one at a time (to avoid exceeding storage limit);
2. JSON data for relearning the final models (intermediate checkpoints were ignored because this eval is the inverse of gradient ascent);
3. a script to plot the eval data;
4. bar charts to illustrate the eval results.

[Davidyz]

Moved to the results repo.