Performance improvements

[rcarver]

A set of performance improvemetns:

• Updates store sparse data; only the modified columns
• Batch replay by grouping insert, update, delete into repeatable queries with IN condition
• Skip duplicate reconcilliation during replay

Internally, we store a simple tab-delimited format instead of real sql string. That makes it easier to parse and aggregate. This is all handled by a type UndoSQL on the table record.

Added a benchmark test. We can't compare to main, but can confirm that the batching mechanism pays off:

┌───────────┬──────┬───────────┬─────────┬─────────┐
│ Operation │ Rows │ Unbatched │ Batched │ Speedup │
├───────────┼──────┼───────────┼─────────┼─────────┤
│ INSERT    │ 2000 │ 123.5 ms  │ 35.2 ms │ 3.5x    │
├───────────┼──────┼───────────┼─────────┼─────────┤
│ UPDATE    │ 2000 │ 307.2 ms  │ 28.0 ms │ 11.0x   │
└───────────┴──────┴───────────┴─────────┴─────────┘