Add the most basic parallelism to the framework, through PyTorch's DDP.
Adjust the dataloaders to also use distributed samplers.

Add other goodies for distributed logging + distributed processing.