v0.8.5 — LeRobotDataset v3.0 export (teach-in / teleop → training-ready data)

LeRobotDataset v3.0 export — an opt-in ⤓ LeRobot button that turns recorded cockpit demos into a LeRobotDataset v3.0 (the Hugging Face / LeRobot standard), so sim demos train ACT / Diffusion Policy / π0 with zero glue.

• Record: each REC start→stop is one episode — the 14 bimanual arm joints sampled at 30 fps (a dense recorder rides the existing teach-in REC hook).
• Export: POST /api/lerobot/export (or the button) writes observation.state / action (ALOHA joint-position convention), per-episode + global stats, tasks and schema into lerobot_datasets/<name>.
• Windows-native & dependency-light: the writer uses only numpy + pandas + pyarrow — no torch / lerobot / ffmpeg — imported lazily, so recording never touches them and the cockpit stays intact if they're missing.
• Validated end-to-end: the format was grounded against real lerobot 0.4.4 on Linux, and a live cockpit recording round-tripped — 2 episodes / 44 frames with real captured motion loaded cleanly with LeRobotDataset(...).

Fully optional and non-breaking. 2 new dep-gated tests (82 headless tests pass).