Capture stateful dataloader state lazily so prepare() does not draw the shuffle permutation

[jayzuccarelli]

What does this PR do?

Fixes #4204

This is the lazy capture I proposed on the issue on Sep 2.

DataLoaderAdapter.__init__ calls base_dataloader.state_dict() right after building the StatefulDataLoader. With num_workers > 0 that builds the iterator and draws the sampler permutation at prepare() time, before synchronize_rng_states runs in __iter__, so each rank shards a different epoch-0 permutation.

Capture it lazily instead: dl_state_dict starts as None and _update_state_dict fills it once iteration starts. state_dict() falls back to the base dataloader if nothing was captured yet, so checkpointing before the first batch works as before.

Added a single-process test (shuffle order follows the seed set after init) and a 2-process num_workers > 0 case in test_distributed_data_loop.py. Both fail on main.

Before submitting

• Was this discussed/approved via a Github issue or the forum? use_stateful_dataloader + num_workers>0 draws epoch-0 permutation during prepare() before cross-rank RNG sync — corrupts DDP data partition #4204
• Did you write any new necessary tests?

Who can review?

@SunMarc