Exit on WAL commit failure

[benbjohnson]

Failures during the last commit phase of WAL operation need to exit as we currently don't have a way to communicate to the SQLite client to rollback the transaction. The commit is detected by the release of the WRITE lock, however, SQLite doesn't handle errors from this syscall so an exit is called instead.

We originally used the WAL write syscall to determine the transaction boundary, however, it is not contiguous when pages are rewritten within a single transaction so it make tracking boundaries inconsistent. We also evaluated checking the fsync syscall, however, this is not used if PRAGMA synchronous=normal is enabled by the client. In the long term, we should re-evaluate the use of write(2) to determine transactional boundaries in WAL mode.