Preserve complete Arrow episodes and their timestep order

[betacatsling]

ArrowStorage consumes one-shot episode indices while building datasets, then tries to iterate them again. It also treats scanner batches as whole episodes, truncating trajectories longer than a batch. Reading complete per-episode tables fixes both issues; serializing writes in order prevents threaded writes from reordering timesteps.

Validation: 4 failed on the unchanged base; 123 passed after the fix, covering Arrow/Parquet long trajectories, one-shot indices, serialization and storage tests. Black/isort and diff checks passed. CPU/local storage only; no GPU, MuJoCo or remote dataset claims. The relevant API documentation is updated.

AI assistance was used for implementation and local verification.

Fixes #351.