Sort out distributed computation

[bpiwowar]

Distributed computation is not working well, and we should switch to DistributedDataParallel for better efficiency

• Samplers should work on independent data subsets
• Checkpointing needs to be done properly

Solve multiple backwards issues:

• Backward is called within trainers (using the no_sync context might lead to problems if the involved parameters are not the same...)
• Micro-batching (using the no_sync context)

See https://pytorch.org/tutorials/intermediate/ddp_tutorial.html

Depends on experimaestro/experimaestro-python#32 since object duplication does not work with the current config/object layout