fix(data): rebuild missing parquet footer (body intact, 125272 frames × 200 episodes)

The original upload was truncated mid-finalisation: leading PAR1 magic and all 6,800 parquet pages (17 cols × 200 row groups) were intact, but the trailing FileMetaData thrift, its 4-byte length, and the closing PAR1 were never written, so pyarrow / the dataset viewer could not parse the footer.

This commit re-adds a synthesised footer that points at the existing page offsets — the first 37,450,342 bytes of the file are byte-identical to the original (SHA-256 verified). pyarrow.parquet.read_table and datasets.Dataset.from_parquet both load all 125,272 rows × 17 columns × 200 unique episodes after the fix.

Recovery scripts: https://github.com/-/this-was-done-locally

data/chunk-000/file-000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3034e5437321fb65abf9a62e9d0b95b6767a512aa2b67b30366be59e2be3829a
-size 37450342
+oid sha256:1f8b71acd7a1e68a9c472061bb404f06a2db70a7d99476a8cbcccdf8f19aefb7
+size 37652523