fix(iterable-dataset): implement stateful protocol for resume

[kcz358]

Problem

MultiModalIterableDataset did not implement state_dict() / load_state_dict(), so StatefulDataLoader had no way to persist the per-worker iteration cursor. On resume, __iter__ unconditionally reset self.cur_idx = 0, so even though global_step and dataloader-level fetch state were restored, the dataset re-walked from the start of the shard — the trainer effectively naive-forwarded a chunk of data before producing the first useful batch.

Fix

• Add state_dict() / load_state_dict() that persist {cur_idx} (called per worker by StatefulDataLoader).
• Guard the cur_idx = 0 resets in both __iter__ branches (packing / non-packing) behind a _resuming flag that load_state_dict sets and __iter__ consumes once.

Per-worker sharding is deterministic given data_seed + world_size + num_workers, so persisting cur_idx alone is sufficient as long as those don't change across resume.

Verification

Two-phase test with scripts/launch/qwen3_vl_test.sh on 4 GPUs / 2 workers per rank, packing enabled, save_steps=2:

Phase	Command	rank0 worker_0 cur_idx	rank0 worker_1 cur_idx
1. Fresh train 2 steps	max_steps=2 save_steps=2	43	24
2. Resume + 2 more steps	max_steps=4 save_steps=2	68 (+25)	66 (+42)

Read directly from checkpoint-*/dataloader_state/*.pt → _snapshot._worker_snapshots[*].dataset_state. Workers continue from the saved position instead of restarting at 0.

Notes

• Packer internal buffer is not persisted, so the very first packed batch after resume may differ slightly from a non-interrupted run. If exact reproducibility is needed, a follow-up can add state_dict to the online packer too.
• No trainer-side changes needed.