Fix stateful dataloader DDP

[SunMarc]

What does this PR do?

Fixes #3938. The tests we had didn't catch the issue that the user was having, so i changed it. Here's the summary of the changes:

• Fix a bug where DataLoaderShard.adjust_state_dict_for_prefetch() incorrectly subtracted num_processes - 1 from state dict counters (_sampler_iter_yielded, _num_yielded) in DDP, causing the resumed dataloader to replay already-consumed batches
• The num_processes - 1 correction in the base class DataLoaderAdapter is only valid for DataLoaderDispatcher (which fetches num_processes batches per step from process 0), not for DataLoaderShard (where each process has
  its own sharded iterator with a single 1-batch look-ahead)
• Strengthen existing test_stateful_dataloader and test_stateful_dataloader_save_state tests: save state earlier (after 3 batches instead of second-to-last), add length assertion, and test both iterable and map-style datasets

Root cause

• In DataLoaderShard.iter, _update_state_dict() is called before the inner next(), so the captured state already equals the number of batches yielded to the user — no DDP adjustment is needed. The base class adjustment of num_processes - 1 caused the resume point to be 1 batch too early, producing duplicate data.

• The bug only affects map-style datasets with use_stateful_dataloader=True in multi-process DDP. Iterable datasets were unaffected because their _sampler_iter_state provides correct resume info independently of _num_yielded. The previous tests didn't catch this because they (a) only used iterable datasets, (b) saved state at the second-to-last batch leaving only 1 batch remaining, and (c) used zip without a length check.