fix: remove spurious routing dependency from JooqDslContextMiddleware - #176

…eated failures

The previous design used a single monitor thread (ZThread.start) that handled
both phase 1 (connection attempt) and phase 2 (disconnect detection after
successful connect). The thread called ZEvent.recv() in blocking mode; closing
the monitor socket from connect() did not unblock the recv(), so the thread
outlived each failed attempt and accumulated across retries.

Root cause: jeromq PAIR/inproc sockets do not unblock a blocking recv() when
the peer socket is closed, so the monitor thread never exited on the failure
path, and ctx.close() in handler.close() was needed to finally terminate it —
causing the ZMQException(ETERM) noise in the CI log and making repeated
connect() calls unreliable under timing pressure.

Fix: split the two phases.
- Phase 1 (connection attempt): removed the monitor entirely. The existing
  CONNECT_TIMEOUT_MS poller loop already handles the failure case correctly.
- Phase 2 (disconnect detection): start a dedicated monitor thread *after* the
  completer handshake succeeds, monitoring only DISCONNECTED/CLOSED events on
  the confirmed connection.

Each monitor thread now has a clean lifecycle: it starts only on success and
exits either on DISCONNECTED (calls close()) or on ETERM (ctx.close()).
No shared monitor socket ownership races remain.