Small fixes from debugging

[jxmorris12]

• Don't always set shuffle=True, set shuffle=shuffle in dataloader_from_dataset (that will always be set to true anyway so far, but this gives us the opportunity to set shuffle=False if we want to)
• Use multiple workers in Dataloaders (this should actually speed up training by a good amount)
• Add model_weights_drop_linear argument, which allows us to re-finetune a model that's already been finetuned once
• Change overloaded batch keyword in train.py to train_batch and test_batch

[jxmorris12]

oh yeah let's merge #18 first and then one of us can fix the conflict by restoring the enumerate() on this branch.