fix: SQLite index compaction dropping incomparable versions

[cpruijsen]

Summary

SQLIndex.compact skips identity advances when grouping rows to move, so values are kept when a version advances to itself under an incomparable frontier (issue #57). The existing compactKey / deleteOldVersionsData statements never insert-then-delete the same primary key.

SQLIndex.compact grouped that identity version with the versions that move onto it, inserted the summed row, then deleted every source version, including the destination. In-memory Index.compact already deleted before rewriting.

Adds the reporter's case to the shared in-memory/SQLite version-index tests.

Provenance: Kevin's reproduction on #57 (in-memory [[10, 3]], SQLite []), matching compaction as meet-of-joins.

Grouping skips identity advances instead of rewriting compact as materialize → delete sources → insert (match in-memory order). Smallest reversible change; same compacted collection as in-memory. Can switch to the rewrite if the SQL path should share control flow with Index.compact.

Fixes #57

Test plan

• packages/d2ts tests/version-index.test.ts: in-memory and sqlite, including "compaction with incomparable versions"
• Confirm the test fails on sqlite (result1 === []) with the source change reverted, and passes with it restored
• Confirm the sqlite case still matches in-memory after a second compact on a later frontier