[Trainer] Add ddp_static_graph option (#45519)
* [Trainer] Add ddp_static_graph option
Expose PyTorch DDP's `static_graph` flag as a new `ddp_static_graph: Optional[bool]`
field on `TrainingArguments`, forwarded through `Trainer._build_accelerator_args`
into Accelerate's `DistributedDataParallelKwargs` (which already supports it).
Completes the set of DDP flags partially exposed today
(`ddp_find_unused_parameters`, `ddp_bucket_cap_mb`, `ddp_broadcast_buffers`).
Defaults to `None`; when unset, the kwarg is never added to `ddp_kwargs`, so
Accelerate's own default (`False`) applies — strictly additive, no existing
behavior changes.
See issue #45518 for full motivation: users with frozen trainable submodules
(e.g. the LLM-frozen head-tuning pattern) today either hit `Expected to have
finished reduction in the prior iteration...` or must pay per-iteration
`find_unused_parameters=True` traversal cost. `static_graph=True` is the
performance-optimal third option that Accelerate/PyTorch already support
but that Trainer couldn't expose.
Tests: positive (True, False) plus regression guard (None must not leak the
kwarg).
Fixes #45518
* Address review: shorten ddp_static_graph docs
