The recorder writes to_timestamp = (n_frames + pts_offset) / fps,
but examples/upload_dataset.py's cleanup_dataset() was overwriting that with
n_rows / fps unconditionally — dropping the pts_offset term. For datasets
captured with NVENC pre-roll (pts_offset > 0), this caused LeRobot's stock
video dataloader to silently truncate the last pts_offset frames (~1.2 s)
of every episode.
Now to_timestamp = n_rows/fps + from_timestamp (per camera), which exactly
recovers the recorder's intended (n_frames + pts_offset) / fps semantics.
n_frames (= parquet length) is unchanged.
Upstream fix in the recorder: trlc-dk1 commit 2100037
