Fix AttributeError when iterating IterableDatasetShard without set_epoch

[vineethsaivs]

What does this PR do?

Iterating an IterableDatasetShard that wraps a dataset carrying a torch.Generator (and no set_epoch method) raises unless the caller happens to call set_epoch first:

class GeneratorIterableDataset(IterableDataset):
    def __init__(self):
        self.generator = torch.Generator()
    def __iter__(self):
        yield from torch.randperm(8, generator=self.generator).tolist()

shard = IterableDatasetShard(GeneratorIterableDataset(), batch_size=2)
list(shard)  # AttributeError: 'IterableDatasetShard' object has no attribute 'epoch'

Root cause: __init__ never initializes self.epoch; the only assignment lives in set_epoch. But __iter__'s first branch reads self.epoch to seed the generator for precisely this dataset shape (hasattr(dataset, "generator") and no set_epoch). Accelerate's own managed path calls set_epoch before iterating, which is why the crash only bites direct users of the class, a usage pattern the existing tests treat as supported (check_iterable_dataset_shards iterates shards directly; it only survives because its test dataset has no generator attribute).

Fix: initialize self.epoch = 0 in __init__, matching SeedableRandomSampler in the same module (and transformers' IterableDatasetShard). Direct iteration now seeds with the default epoch 0; set_epoch semantics are unchanged.

Test: test_iterable_dataset_shard_without_set_epoch reproduces the exact shape (generator-carrying dataset, no set_epoch call): fails before with the AttributeError at data_loader.py:346, passes after; also asserts the default seeding is deterministic. Full tests/test_data_loader.py: 25 passed, 13 skipped. ruff check / ruff format --check clean on both files.

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests?