Implement WAL group commits

[adsharma]

Summary

• add WAL group commit sequencing so multiple committers can share one durable sync
• split transaction commit into durable WAL append followed by in-memory publish
• preserve publish order and single-writer behavior while allowing committers to wait during sync

Benchmark

C++ benchmark source: node_write_bench.cpp. Optimal was 64 threads at 6k writes/sec.

Setup:

• on-disk temp database under /private/tmp
• enableMultiWrites=true
• autoCheckpoint=false
• forceCheckpointOnClose=false
• one autocommitted CREATE (:person {id, name}) per write
• one Connection per writer thread
• count verified after each run

1,000 writes per thread

Branch	Threads	Total writes	Writes/sec
main	1	1,000	249.068
group commit	1	1,000	244.118
main	2	2,000	256.391
group commit	2	2,000	248.242
main	4	4,000	254.625
group commit	4	4,000	495.887
main	8	8,000	258.332
group commit	8	8,000	951.349

2,000 writes per thread

Branch	Threads	Total writes	Writes/sec	Speedup
main	4	8,000	265.451	1.00x
group commit	4	8,000	493.874	1.86x
main	8	16,000	283.718	1.00x
group commit	8	16,000	1,016.86	3.58x
main	16	32,000	290.376	1.00x
group commit	16	32,000	2,115.63	7.29x

Main remains fsync-bound at roughly 250-290 writes/sec as writer concurrency increases. With group commits, throughput scales to roughly 1,017 writes/sec at 8 writers and 2,116 writes/sec at 16 writers for this benchmark.

[adsharma]

Cc: @loganpowell

[adsharma]

• WAL sync failures now still expose the allocated commit sequence to TM cleanup.
• Exception cleanup now clears active write transaction state and wakes waiters.
• Direct Transaction::commit() / publishCommit() now guard invalid commit timestamps.
• Replaced the begin-transaction sleep loop with a condition variable.
• Fixed WAL::reset() ordering.

[adsharma]

Postgres and SQLite handle fsync() failures differently. Postgres poisons the WAL and blocks any future commits. SQLite continues with an "uncertain durability" state.

We're taking the Postgres approach.