Preserve pretraining epoch across resume - #2313

What changed

• make CycleIterator save and restore its epoch through Fabric checkpoints
• create the pretraining iterator before checkpoint loading and reuse it in fit
• keep old checkpoints compatible by loading the existing state first, then restoring the optional iterator state from Fabric's remainder
• cover mid-epoch and exact epoch-boundary resume with a real LitData StreamingDataLoader

Why

The current litdata==0.2.59 pin already restores the dataloader cursor correctly: after a mid-epoch checkpoint it yields the remaining tail once and then starts a complete new epoch. However, pretrain still constructs CycleIterator after loading and resets its epoch counter to zero.

This addresses the remaining epoch-tracking problem described in #1712 without changing dataloader cursor semantics. The boundary handling is needed because a restored StreamingDataLoader can begin the next epoch itself before CycleIterator observes StopIteration.

Validation

• unmodified baseline: the real Fabric checkpoint regression fails because CycleIterator contains an unpicklable generator, and the pretraining resume path has no iterator state
• pytest tests/test_pretrain.py tests/test_utils.py -q: 59 passed, 9 skipped
• all applicable changed-file pre-commit hooks passed, including Ruff check and format
• git diff --check passed