fix: batch WAL frame writes into one write_at per commit

[iheitlager]

Summary

• WalWriter::append_frame issued its own write_at syscall per dirty page instead of batching a transaction's frames into a single write — an O(n) syscall pattern in the hot commit path.
• Accumulate frames into a pending buffer during append_frame; sync() now issues one write_at covering the whole run, then fsyncs exactly once (ADR-0026's per-commit rescan behavior is unchanged — only the writes feeding that one fsync are batched).

Profiling (#635)

update_batch_tx_wal benchmark: sqlite-rs 27-32ms vs C SQLite oracle ~4.3ms on bench_1mb.db.

Instrumented the commit path directly (not via a spawned process, to avoid process-startup noise): for this benchmark's 34-page commit, open_existing rescan is ~6-7ms and the batched write+fsync is ~4.5-5ms — roughly ~11ms of the ~27-32ms total. The remaining ~15-20ms is VDBE table-scan interpretation overhead for the unindexed WHERE bucket = 5 clause (~1µs/row × 16,700 rows) — not a WAL-layer cost, and not something this PR attempts to fix.

This fix is a real, tested improvement (removes N syscalls → 1 for multi-page commits) but does not close the 7x gap for this specific benchmark on its own. Filed follow-ups for the remaining costs:

• perf: VDBE table-scan interpretation overhead on unindexed WHERE (~1µs/row) #639 — VDBE table-scan interpretation overhead
• perf: WAL commit re-reads and rescans the entire -wal file every flush (ADR-0026 debt) #640 — ADR-0026's flagged per-commit WAL rescan cost

spend: within estimate (single profiling-driven fix, no scope creep)

Closes #635

Test plan

• cargo test --lib — 953 passed
• cargo test --test tier0 --test tier1 --test tier2 --test tier3 — all pass, including t3_wal_writing_and_live_interop
• cargo clippy --lib -- -D warnings — clean
• A/B benchmark: confirmed baseline (pre-fix) and post-fix update_batch_tx_wal timings, batching alone doesn't move this particular benchmark much (page count too small — 34 pages — for syscall count to dominate), consistent with the profiling above