Combine / rename dist.ddp and dist.spmd into dist.torchrun

[schmidt-ai]

Description

Currently, dist.ddp and dist.spmd are basically identical (the latter being a lightweight wrapper on the former). Also, they could be named more explicitly — dist.ddp doesn't actually involve Distributed Data Parallel, it just calls torchrun.

Motivation/Background

All else equal, simplification and explicit naming are good. For example, users leveraging Fully Sharded Data Parallel instead of DDP may find it confusing that they should be using dist.ddp.

Detailed Proposal

Refactor components/dist.py by combining the methods for ddp and spmd into one method called torchrun. Update docs, tests, examples, and callsites as appropriate.

Alternatives

1. Leave thing as-is.
2. Remove ddp by rolling it into spmd and keep the spmd method, so dist.spmd is the only available command and it has a "good enough" name.

Additional context/links

@danielbear