fix(record): remove the empty dataset dir a failed recording leaves behind

[ravediamond]

Closes #112.

When a recording fails before any episode is saved, the dataset directory it already created gets left on disk and shows up in the dataset list as if something was recorded.

Added _cleanup_failed_recording(), called from the recording worker's error handler: if the session wasn't resuming an existing dataset and saved 0 episodes, it removes the dataset directory. Best-effort — logs and swallows any cleanup failure rather than masking the original recording error.

Also pulled the path-traversal guard out of handle_delete_dataset into _resolve_dataset_dir so the new cleanup path reuses it instead of a third copy, and switched it to read the cache root from the environment at call time (same as episode_media.lerobot_cache_root()) instead of importing lerobot's HF_LEROBOT_HOME, which is frozen at import time — the old version would have silently used the wrong root if the env var changed after the module first loaded. Writing a test for it is what surfaced this.

Tested: pytest (277 passed, 9 new tests covering the traversal guard, delete-dataset, and the four cleanup branches), ruff check/format clean. No frontend changes.