Parquet files + a JSON manifest

A table format is just metadata that points at data files. In this stage you split the two: row data goes into immutable Parquet files; a tiny JSON manifest records which files make up the table.

You edit tableformat.py. The provided objectstore.py is a key→bytes store over a local directory — treat it like S3. Do not edit objectstore.py.

Implement

• Table.create(store, name, schema) -> Table — register a brand-new, empty table. Fail if it already exists. Write the first metadata object at key "<name>/metadata/v1.json".
• Table.load(store, name) -> Table — open an existing table (raise if missing).
• table.append(arrow_table) -> snapshot_id — write the rows to a new Parquet file under "<name>/data/<something>.parquet", then record that file in a new metadata object. Return the new snapshot id (an int).
• table.scan() -> pyarrow.Table — read all data files listed in the current metadata and return their concatenation as a single pyarrow.Table.

Rules that make it a format, not a heap of files

• Data and metadata are separate objects. The manifest is JSON and lists data files by key, with at least path and num_rows. The tests open your JSON manifest and your Parquet files directly with a real reader.
• Data files are immutable. append never rewrites an existing file; it adds a new one.
• Each metadata object is a new version v<N>.json; the "current" table is the highest version present.

API you are starting (later stages extend the same file)

Table.create(store, name, schema) -> Table
Table.load(store, name) -> Table
table.append(arrow_table) -> snapshot_id
table.scan(filter=None, snapshot_id=None) -> pyarrow.Table
table.snapshots() -> list[Snapshot]      # stage 2
table.delete(predicate) -> snapshot_id    # stage 5
table.compact() -> snapshot_id            # stage 7
table.expire(older_than_ms) -> int        # stage 7

Run de test to check this stage. Later stages re-run these tests as a regression suite, so keep append/scan working.