What this fixes

Iterating a prepared HF datasets.IterableDataset for more than one epoch crashes
with DataSourcesShufflingDisallowed when the dataset was built via .skip() /
.take() (#4080).

DataLoaderShard.__iter__ and DataLoaderDispatcher.__iter__ call
self.set_epoch(self.iteration) at the start of every pass, forwarding a nonzero
epoch to the wrapped dataset. On the second epoch the dataset tries to reshuffle its
data sources at iteration time, which .skip()/.take() iterables forbid, so it
raises DataSourcesShufflingDisallowed.

The fix

Catch DataSourcesShufflingDisallowed while iterating and reset the dataset's epoch
to 0, then re-create the iterator. These datasets can't reshuffle their sources
between epochs anyway, so this lets them iterate across multiple epochs without
changing behaviour for datasets that do support per-epoch reshuffling. The exception
is imported defensively so datasets stays optional.

Repro (from the issue)

  import datasets, accelerate, torch.utils.data
  ds = datasets.Dataset.from_dict({"a": list(range(20))}).to_iterable_dataset().skip(10)
  dl = accelerate.Accelerator().prepare_data_loader(torch.utils.data.DataLoader(ds, batch_size=2))
  for b in dl: pass
  for b in dl: pass   # raised DataSourcesShufflingDisallowed before this fix

Tests

Adds test_iterable_dataset_blocked_source_shuffling_multiple_epochs, covering both
the dispatch and non-dispatch paths. It fails on main and passes with the fix; the
rest of tests/test_data_loader.py is unaffected.

Fixes #4080