Refactor loading of data to use a data loader

I would suggest wrapping the data loading into a pytorch data loader and just a sampler for shuffling and batching.

One needs to

1. Implement a dataset that takes care of the loading tokenization and returns instances. I think a map-style dataset that returns dict/object (like SentenceEvidence)
2. Implement a collate_fn that takes a list of these object and returns a batch
3. Then one can just use:

def collate_and_padd_batch(...):
   ...
dataset = EraserDataset(..., split='train')

loader = DataLoader(dataset, batch_size=1, shuffle=True, num_workers=0, collate_fn=collate_and_padd_batch)

for batch in loader:
  # to training
  ...

For more information see