Remove blocking IO from mvcc bootstrap and recovery paths

[PThorpe92]

Lift blocking IO part 2: schema reparse + MVCC bootstrap/recovery

Removes synchronous io.block/wait_for_completion from the MVCC open path so it no longer hangs on backends without a synchronous IO pump (e.g. WASM, where io.step() is a no-op). The whole bootstrap → checkpoint-reconcile → schema
reparse → metadata-table init → logical-log replay sequence now yields IO cooperatively via IOResult instead of blocking.

Major changes

Non-blocking statement runner

• Statement::run_ignore_rows_nonblock / run_with_row_callback_nonblock:
  drive a statement to completion returning IOResult, surfacing the pending
  completion (via take_io_completions) instead of pumping io.step(). The
  VDBE interpreter already yields StepResult::IO; these just bridge it to the
  IOResult model. No interpreter changes.

Schema reparse

• reparse_schema / reparse_schema_with_cookie lifted to a ReparseSchemaState
  machine (read cookie → scan sqlite_schema → load custom types → refresh
  stats), carrying the half-built schema + captured table-valued functions +
  reparse guard across yields.
• parse_schema_rows, refresh_analyze_stats, and read_current_schema_cookie
  lifted to IOResult accordingly.

MVCC bootstrap

• bootstrap_nonblock reworked into a linear state machine that drives, all
  non-blocking: interrupted-checkpoint reconciliation, schema reparse, the
  persistent_tx_ts_max read, metadata-table create/seed, and the metadata IO
  chain. Dead blocking shims removed.

Logical-log recovery

• maybe_recover_logical_log lifted to a RecoverLogicalLogState machine. Setup
  phases (header / tx-ts / cookie / sqlite_schema scan) yield IO; the replay
  loop carries its accumulators in RecoverCtx across per-frame next_frame
  yields. The per-frame replay body is unchanged in behavior. Driven from a new
  bootstrap Recover phase.

Checkpoint hang fix (mvcc/database/checkpoint_state_machine.rs path)

• maybe_complete_interrupted_checkpoint's DriveEarlyTruncate was recreating
  its CheckpointResult on every re-entry, so truncate_wal (which tracks
  truncate/sync progress through that struct) never completed and spun under
  io.block. The result now persists in the state variant.

[PThorpe92]

Will kick off an Antithesis run from this branch once the CI is green.

Whopper seems to have surfaced an issue here with recovery, fixing

Will address the perf regression as best I can too

[codspeed-hq]

Merging this PR will not alter performance

✅ 638 untouched benchmarks
⏩ 105 skipped benchmarks¹

---

_{Comparing PThorpe92:nonblock (8551c0d) with main (c70a032)}

Footnotes

1. 105 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[PThorpe92]

note to self: looks to be that this function is not safely re-entrant

it was re-entrant it was just stupid because it re-parsed the whole frame if any frame yielded mid-way, so a real state machine was added