The driver's _handle_finished calls _join_test_process inline, which
blocks the single-threaded ZMQ REP loop for up to 30s. During that
block, clients retry their FINISHED messages, which queue up on the
REP socket. When the driver unblocks, duplicate FINISHED messages
cause a KeyError on `del self.active_tests[test_key]` that kills the
entire session.

Changes:

- Add idempotency guard in _handle_finished: check active_tests and
  finished_tests before processing. If the test was already finished,
  log elapsed time since original event_time. Return early instead of
  crashing with KeyError.

- Track last 100 received messages in a deque ring buffer. On fatal
  exception in the main event loop, dump the recent message history
  (test_key, event_type, event_id, message_id).

- Add FATAL_ERROR_TAG to the exception handler log line.

- Log a warning when event processing exceeds SLOW_EVENT_THRESHOLD_S
  (5s), using try/finally around _handle() so it fires on both
  success and exception paths.

- Increase Sender.REQUEST_TIMEOUT_MS from 3s to 35s so the first
  send attempt outlasts the 30s join timeout. This is needed because
  non-FINISHED events are not idempotent and cannot safely be retried.

Allow overriding REQUEST_TIMEOUT_MS via constructor param, defaulting
to the class constant (35s). This lets the check_timeout test pass a
small timeout (100ms) so the spawned Sender subprocess finishes in
~0.5s instead of ~175s (5 retries * 35s).