Feature/ray train update

[davidegraff]

Description

this PR updates the distributed training code to bring it up-to-date with ray 1.11.

Summary:

• rename mpnn.sgd module to mpnn.ray
• update all code in the module to move from ray.sgd.v2 to the new ray.train interface
• updated some MoleculeDataset and DataLoader code to return arrays/tensors rather than lists (to work better with training loops overall)
• updated training/validation loop code to use pytorch tensor functions rather than rebuilding tensors from scratch. E.g., from

  mask = torch.tensor([[bool(y) for y in ys] for ys in targets], device=device) # <-- necessary when targets was List[List[Optional[float]]

  to

  mask = ~torch.isnan(targets) # <-- possible because targets is now torch.Tensor

• formatted and linted the entire molpal.models module

Checklist

• [x] linted with flake8?
• [x] (if appropriate) unit tests added?
• [x] ready to go?