fix(datasets): address sampler review (batch_size resume guard + docs)

- Record batch_size in training_step.json alongside num_processes and feed
  the checkpoint's value into compute_sampler_state on resume; warn when it
  differs (per-rank sample-exactness needs the same batch size).
- Document the set_epoch vs __iter__ auto-advance coupling on EpisodeAwareSampler
  (callers should rely on exactly one mechanism per run).
- Note the broadened (reproducibility-breaking) sampler guard and the no-generator
  distributed sharding correctness in lerobot_train.py.
- Add load_training_batch_size + parallel tests.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/utils/test_train_utils.py

@@ -20,6 +20,7 @@ from unittest.mock import Mock, patch
 from lerobot.common.train_utils import (
     get_step_checkpoint_dir,
     get_step_identifier,
+    load_training_batch_size,
     load_training_num_processes,
     load_training_state,
     load_training_step,
@@ -75,6 +76,17 @@ def test_load_training_num_processes_absent_returns_none(tmp_path, optimizer, sc
     assert load_training_num_processes(tmp_path) is None
 
 
+def test_save_training_state_records_batch_size(tmp_path, optimizer, scheduler):
+    save_training_state(tmp_path, 10, optimizer, scheduler, batch_size=32)
+    assert load_training_batch_size(tmp_path) == 32
+
+
+def test_load_training_batch_size_absent_returns_none(tmp_path, optimizer, scheduler):
+    # Checkpoints written before the batch size was recorded must still load (back-compat).
+    save_training_state(tmp_path, 10, optimizer, scheduler)
+    assert load_training_batch_size(tmp_path) is None
+
+
 def test_update_last_checkpoint(tmp_path):
     checkpoint = tmp_path / "0005"
     checkpoint.mkdir()