`table.slice(...).to_pandas()` produces pandas ExtensionArrays for
`array[float32]` columns (e.g. `observation.states.end.orientation`)
on newer pandas/pyarrow combos, which then fail in
`pa.Table.from_pandas` inside `Dataset.from_pandas(...).to_parquet(...)`.
Skip the pandas round-trip and wrap the `pa.Table` slice in a
`Dataset` directly with `Dataset(episode_table).to_parquet(...)`.
This preserves the HuggingFace dataset metadata that `Dataset.to_parquet`
writes, while avoiding the ExtensionArray crash. No version pin on
datasets/pyarrow needed.
Closes#87
Qizhi Chen