fix(pglite-socket): recover query queue after an internal execution error

[Brown-Sage]

Description

Fixes the queue deadlock described in #1046 (Defect B).

In QueryQueueManager.processQueue(), a single throw from execProtocolRawStream (e.g. a WASM abort or a race with db.close()) hit the catch block which called query.reject() and then returned — skipping this.processing = false. Since enqueue() only kicks the loop if (!this.processing), one internal failure permanently deadlocked the queue: every subsequent query from any connection was queued and never executed.

Changes

• Reset processing in a finally block so the flag is always restored
• continue draining the queue after a failed execution instead of returning, so one bad message doesn't kill the remaining queued queries
• Transaction affinity semantics are intentionally unchanged (that's Facet C in [BUG]: pglite-socket: concurrent extended-protocol batches interleave across connections; one internal throw permanently deadlocks the query queue #1046 and is separate design discussion)

Regression tests

New tests/queue-recovery.test.ts, modeled on the fault-injection repro from #1046:

1. Injects a single rejection into db.execProtocolRawStream on the first Parse message, asserts the failing client gets an error and a fresh connection is still served afterward
2. Same injection with two concurrent connections; client B must complete regardless of A hitting the injected failure

Both tests fail against unfixed main (queries hang) and pass with this patch.

Also adds a changeset for @electric-sql/pglite-socket.

Fixes #1046