Create a more cohesive training/evaluation interface

[amanchoudhri]

Currently there are a number of different ways to train models (run_gpu.sh, run_batch.sh, end_to_end.sh, requesting a gpu manually and using main.py, etc), along with assess and eval (assess.py, run_eval.sh). Some of these scripts no longer work, and it's not clear which are the preferred / best ways to train models.

Recommendation:

• one preferred way to train a single model, with options to assess it after training or finetune it.
• one preferred way to train a group of models in parallel
• one preferred way to assess an already-trained model on a specified dataset