Loader checkpoint restoration restarts before a None item

[aswanth-07]

Bug

torchdata.nodes.Loader cannot restore a checkpoint whose next item is None when restart_on_stop_iteration=True (the default). The look-ahead in LoaderIterator.has_next() treats None as meaning that no item was cached, decides the restored iterator is exhausted, and restarts it from the beginning.

Reproduction

from torchdata.nodes import IterableWrapper, Loader

loader = Loader(IterableWrapper([1, None, 2]))
iterator = iter(loader)
assert next(iterator) == 1
state = loader.state_dict()

restored = Loader(IterableWrapper([1, None, 2]))
restored.load_state_dict(state)
print(list(restored))

On current main (e640e6f), this prints:

[1, None, 2]

The expected continuation is:

[None, 2]

The restart duplicates data that was consumed before the checkpoint. Calling LoaderIterator.has_next() directly also returns False and consumes the value when the next item is None.

Root cause and proposed fix

LoaderIterator._cached_item currently uses None both as a valid data value and as its empty-cache sentinel. Tracking cache occupancy separately avoids that collision without changing the public API or the look-ahead behavior for other values.

I have a focused regression test and a small fix ready.

[aswanth-07]

/assign