Multiple Eval Sets MNLI

[benja-matic]

• Created a mixin where trainer.evaluate() loops over a list of evaluation sets
• Refactored and fixed various checks, including new checks to ensure the above mixin is used when finetuning mnli
• Making a good default behavior for handling mnli required several adjustments, including new run_args and updates to task_hyperparams updates in run.py
• Added configs for numerous experiments
• Removed some configs that are no longer in use to avoid confusion, added labels
• Gathered hyperparameters for small berts
• Added hyperparam analysis code so that hps for one task serve as a proxy for another
• Updated the prediction and link predictions components of run_utils to handle multiple test sets