…vered bugs

…er retry]

xWrite in the v2 VFS was calling resolve_pages for every dirty page,
even when the write was a full-page aligned overwrite. For page numbers
beyond db_size_pages (newly allocated pages), this meant fetching from
the engine to get data that doesn't exist.

A 256-row INSERT transaction (~1MB of dirty data) was making 288 round
trips to the engine: one get_pages RTT per new page allocation. At
128ms staging RTT that's ~37s of theoretical network time vs the ~130ms
needed for just the final commit.

Now:
- Full-page aligned writes (offset % page_size == 0 && amt % page_size == 0)
  skip resolve_pages entirely and use a zero-filled page as the base.
- Partial writes filter out pages > db_size_pages before calling
  resolve_pages, synthesizing zero pages locally for new allocations.

Direct engine profiling tests verify 0 engine fetches for 1MB, 5MB, and
100-hot-row-updates workloads.

Adds new stress-test workloads to the kitchen-sink bench:
churnInsertDelete, mixedOltpLarge, growingAggregation,
indexCreationOnLargeTable, bulkUpdate1000Rows, truncateAndRegrow,
manySmallTables.