StatefulDataLoader loses alignment on second resume within an epoch

[mmshad]

StatefulDataLoader misaligns the data order on the second (and later)
save/load cycles within a single epoch:

• __iter__ sets self._batches_yielded = 0 on every call.
• load_state_dict reads batches_yielded into a local variable and
  never writes it back onto self.
• The sampler exposes a single-shot set_skip that is consumed the first
  time iter() advances.

Sequence that breaks:

1. Run 1 yields 30 batches, state_dict() records batches_yielded=30.
2. Run 2 load_state_dict applies sampler.set_skip(30*bs). It then calls
   iter(), which consumes the sampler skip, and _batches_yielded is
   reset to 0.
3. Run 2 yields 10 more batches. state_dict() records
   batches_yielded=10, not 40. The sampler state also only reflects the
   10-batch delta because it was re-seeded at iter().
4. Run 3 load_state_dict applies sampler.set_skip(10*bs). The loader
   skips 10 batches, not 40.

On busy clusters with multiple preemptions per epoch, this silently shifts
the data order on every resume after the first.

Fix

• load_state_dict writes batches_yielded to self._batches_yielded.
• __iter__ re-applies sampler.set_skip(self._batches_yielded * self.batch_size) and no longer zeros the counter. _batches_yielded
  now advances monotonically within the epoch and resets only when
  StopIteration fires.

Coverage

tests/integration/test_data_resumption.py::test_double_resume_within_same_epoch:
three-way resume (30 -> 40 -> ground-truth batches 40..49) asserts the
batches on the second resume match a single continuous run batch-for-batch.