fix(meta/episodes): regenerate per-episode index parquet

The original meta/episodes/chunk-000/file-000.parquet on the Hub had the same truncated-footer corruption as the data parquet (PAR1 head, no PAR1 tail; pages intact). After the data parquet was repaired (previous commit on this branch) the per-episode index parquet was regenerated from scratch by:

1. Grouping the data parquet rows by episode_index (200 episodes, 125,272 frames total).
2. Walking the videos/<key>/chunk-000/file-*.mp4 files to compute per-camera (chunk_index, file_index, from_timestamp, to_timestamp) for every episode (matches LeRobot's writer roll-over rule based on data_files_size_in_mb).
3. Computing per-episode stats {min,max,mean,std,count,q01,q10,q50,q90,q99} for every numeric and video feature using the same RunningQuantileStats algorithm as lerobot.datasets.compute_stats (a single deterministic ~125-frame sample per episode for video features, normalised to [0,1] with shape (3,1,1)).
4. Materialising the 165-column schema (episode_index, tasks, length, data/{chunk,file}_index, videos/<key>/{chunk,file,from_timestamp,to_timestamp}, stats.<feature>.<stat>, meta/episodes/{chunk,file}_index, dataset_{from,to}_index) and writing it as a single SNAPPY-compressed row group.

The recovered file passes pyarrow.parquet.read_metadata + read_table and datasets.Dataset.from_parquet end-to-end. dataset_to_index of the last episode (125272) matches the data parquet row count exactly.

meta/episodes/chunk-000/file-000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5ae4fd95bd3dcc9b4592b1eeae8ec9f1218d3303353228bae99f2ceb7a9c642d
-size 1324150
+oid sha256:cfd5eac2c4b5091d6045752295bfcb8cae4684b1cd904fd39782e32d513ce87f
+size 872173