fix: placeholder stats for deferred video keys respect channel axis

[shuheng-liu]

What this does

LeRobotDataset.save_episode synthesizes placeholder statistics for deferred video keys so downstream stat-aggregation code never sees a missing entry. The fallback pulled the channel count from shape[0], which is only correct for (C, H, W) callers — LeRobot v2.1's default image convention is (H, W, C) with names ["height", "width", "channel"], where shape[0] is the height.

Concretely, a 480×640×3 deferred video key produced placeholder stats with shape (480, 1, 1). On the second save_episode call, aggregate_stats → _assert_type_and_shape rejected it because the assertion requires (3, 1, 1) for any feature with "image" in the key. Datasets with only one episode never hit this because aggregate_stats needs two entries to merge.

Fix: look up the channel axis via the feature's names list (handling both "channel" and "channels"), falling back to the existing shape[0] behavior when names are absent or don't mention a channel axis, and to 3 as a last resort. Preserves current behavior for every (C, H, W) caller, including the existing tests at tests/datasets/test_datasets.py that declare shape (3, 96, 128) with names ["channels", ...].

Adds a regression test that declares a deferred video key in (H, W, C) order and saves two episodes — this fails on main and passes with the fix.

How it was tested

Ran a test that declares a deferred video key in (H, W, C) order and saves two episodes — this fails on main and passes with the fix.

How to checkout & try? (for the reviewer)

Run a test that declares a deferred video key in (H, W, C) order and saves two episodes — this fails on main and passes with the fix.

Checklist

• I have added Google-style docstrings to important functions and ensured function parameters are typed.
• My PR includes policy-related changes.
  • If the above is checked: I have run the GPU pytests (pytest -m "gpu") and regression tests.