What does this PR do?

Fixes #21567

Problem

When using Fabric with DDP and gradient accumulation, the following warning
was emitted on every backward pass:

UserWarning: The AccumulateGrad node's stream does not match the stream of the
node that produced the incoming gradient. This may incur unnecessary
synchronization and break CUDA graph capture if the AccumulateGrad node's
stream is the default stream...

This warning did not appear when using plain PyTorch DDP directly.

Root Cause

The original setup_module initialized DistributedDataParallel inside a
new side-stream context (torch.cuda.Stream()):

ctx = torch.cuda.stream(torch.cuda.Stream())
with ctx:
    return DistributedDataParallel(...)

This was intentional for supporting CUDA graph whole-network capture, which
requires DDP to be initialized on a side-stream (see PyTorch docs:
https://docs.pytorch.org/docs/2.12/notes/cuda.html#id5).

However, for normal training (the vast majority of use cases), this causes a
stream mismatch: DDP registers its AccumulateGrad hooks on the side-stream
during initialization, but all subsequent forward/backward passes run on the
default stream. PyTorch detects this cross-stream node reference and emits
the warning.

Plain PyTorch DDP does not hit this because users initialize DDP directly
without any stream context, it defaults to the default stream, so there is
no mismatch.

Fix

Detect whether we are currently inside a CUDA graph capture context using
torch.cuda.is_current_stream_capturing(), and pick the appropriate stream:

• Normal training (capturing=False): initialize DDP on the default
  stream. No mismatch, no warning.
• CUDA graph capture (capturing=True): initialize DDP on a new
  side-stream as before (required by PyTorch). Additionally suppress the
  AccumulateGrad warning globally since the mismatch is intentional in
  this context.

This fix is applied to both DDPStrategy in lightning_fabric and
DDPStrategy in pytorch_lightning.

Testing

Verified with the reproduction script from #21567 across 4 GPUs (torch 2.11
and 2.12). Warning no longer appears under normal DDP + gradient accumulation
training.

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list: