Added DistributedStrategy interface with support for DDP

[tgaddair]

Closes #2886.

Usage:

backend:
  trainer:
    strategy: ddp

This will allow us to use PyTorch 2.0 model compilation with distributed training:

https://pytorch.org/get-started/pytorch-2.0/

Caveats:

• Will hang if an exception raised in only one of the training workers due to this issue in Ray Train. Should be addressed by the upgrade to Ray 2.1 / 2.2 in Support Distributed Training And Ray Tune with Ray 2.1 #2709.
• Requires that all checkpoints are readable from every worker, as DDP does not support Horovod's broadcast state operations, which allow the files to only exist on the coordinator (should not be an issue for Ray users, as they already have to work with the assumption that the coordinator could land on any node).

[github-actions]

Unit Test Results

         6 files  ±0           6 suites  ±0   4h 15m 57s ⏱️ + 39m 29s
  3 568 tests +2    3 496 ✔️ +2    72 💤 ±0  0 ❌ ±0 
10 704 runs  +6  10 471 ✔️ +6  233 💤 ±0  0 ❌ ±0 

Results for commit b99e580. ± Comparison against base commit b320bec.

♻️ This comment has been updated with latest results.

[justinxzhao]

LGTM. Strategy seems like a nice abstraction to add support for pytorch DDP

[justinxzhao]

Should self._distributed be an attribute of the Backend class?

[tgaddair]

Only for Horovod. In the case of Ray, we are not in the right context from the backend to have access to the distributed api.