mvcc: fix autoincrement handling

[PThorpe92]

Description

Addressing #5687 and adds some of the tests added in that PR.

Problem

AUTOINCREMENT in MVCC had two correctness issues:

1. Concurrent writers could fail with WriteWriteConflict on sqlite_sequence.
2. In mixed INSERT / UPDATE rowid / DELETE workloads, auto-assigned rowids could diverge from SQLite semantics (including monotonicity failures in fuzzed cases).

Solution

At a high level:
MVCC now treats sqlite_sequence specially for conflict/read semantics:

• ignore row-level WW conflicts on sqlite_sequence at delete/commit time
• when reading sqlite_sequence, prefer the visible version with the highest seq

Insert translation was tightened so sequence-row updates are stable and deterministic in MVCC:

• MVCC uses fixed sequence-row identity strategy where appropriate
• update path uses REQUIRE_SEEK to avoid duplicate sequence-row artifacts
• AUTOINCREMENT allocation now uses a floor in MVCC:
  • floor = max(sqlite_sequence.seq, current table max rowid)
  • that floor is passed into NewRowid so concurrent allocation remains allocator-driven but never falls behind sequence/table state
• Rowid-update paths no longer poison allocator state:
  • UPDATE ... SET rowid=... writes are marked and applied without bumping the allocator as if they were normal autogen inserts
• Negative rowid edge handling in NewRowid initialization is normalized to SQLite-compatible behavior for autogen allocation.

[turso-bot]

Please review @sivukhin