feat: add record-mode to Store for DPO/GRPO

- Store gains fetch_record/num_records alongside stream fetch/__len__
- save_bin/load_bin support per-record offsets via record_keys param
- H5Store/MmapStore/JsonlStore all support dual stream+record access
- DPODataset/GRPODataset use fetch_record, no cross-record concat
- dpo_collate_fn + collate_fn wired through TrainConfig
- fixes attention context leakage in DPO from windowed concatenation

astrai/dataset/sampler.py

@@ -5,7 +5,15 @@ import torch.distributed as dist
 from torch.utils.data import Dataset, Sampler
 
 
-class ResumableDistributedSampler(Sampler[int]):
+class RDSampler(Sampler[int]):
+    """Resumable Distributed Sampler.
+
+    A distributed sampler that supports checkpoint-based resume: iteration
+    state (epoch, position) is tracked so training can continue from the
+    exact sample after a restart.  Shards the dataset across
+    ``dist.world_size`` replicas with optional shuffling.
+    """
+
     def __init__(
         self,
         data_source: Dataset,