Set WAL journal mode and synchronous=NORMAL: ~4x on the write path #3

SqliteStore opens the database without setting any journal or synchronous pragma, so it runs on SQLite's defaults: journal_mode=DELETE and synchronous=FULL. Every write does a rollback-journal round trip and a full fsync.

Measured per put():

default (DELETE + FULL)        0.078 ms
WAL + synchronous=NORMAL       0.019 ms

Roughly 4x, and it lands on the publish hot path — storeRetained, outgoingEnqueue and incomingStorePacket all go through put(), so it is paid per message rather than per connection.

WAL also lets readers run concurrently with a writer, which suits a broker workload where retained/outgoing scans overlap publishes.

Fix

Set the pragmas at open:

PRAGMA journal_mode = WAL;
PRAGMA synchronous = NORMAL;

synchronous=NORMAL under WAL is the usual pairing and is durable across process crashes, trading only the last transactions on an OS crash or power loss. That is a judgement call for a broker persisting in-flight QoS state, so it is worth deciding deliberately rather than inheriting the default — and worth exposing as a constructor option if the safer setting is wanted by default.

Note WAL does not work on a database on a network filesystem, and it creates -wal/-shm sidecar files, so anything that copies or backs up the store file needs to know.

Found while reviewing #1 and left out of it: repo-wide performance configuration, unrelated to that PR's security fix.