Compaction + snapshot expiration

Many small appends and delete files make reads slow. Maintenance fixes that without breaking readers pinned to old snapshots.

Implement

• table.compact() -> snapshot_id — read all live rows of the current snapshot (data minus deletes), write them into one new data file, and commit a snapshot whose manifest is that single file with no delete files. Do not physically remove the old data/delete files yet — a reader time-traveling to an earlier snapshot still needs them.
• table.expire(older_than_ms) -> int — drop snapshots whose timestamp_ms is < older_than_ms (always keep the current one), then garbage-collect any data or delete file that no retained snapshot references. Return the number of files physically deleted.

The invariants the tests enforce

• After compact(), the current scan is unchanged (same rows) but the current manifest is a single file with no deletes.
• A reader pinned to a pre-compaction snapshot still reads the correct old rows — because compaction only adds; it never deletes the old files.
• expire must never delete a file still referenced by a snapshot it kept. Expiring old snapshots first makes their now-unreferenced files collectible; the live snapshot's files are always retained.

Keep stages 1–6 green.