Fix resilience relay bind: drain relay poll-waiters on stop/kill

[ksuderman]

Refs #469.

The Resilience Suite intermittently fails at fixture setup with
TimeoutError: Pulsar did not bind relay consumers within 60.0s
(test/resilience/harness/pulsar_control.py).

Cause

For AMQP modes, PulsarControl.kill() force-drops the stale broker consumer via
_force_drop_setup_consumer_connections() so the next setup can't race a lingering
registration. Relay mode had no equivalent. Pulsar's relay control consumer holds a
30 s long_poll waiter on the relay (pulsar/messaging/bind_relay.py); after a
stop/kill that waiter lingers on the relay until it notices the dropped connection.
Because the readiness check _relay_has_pulsar_setup_waiter() returns True on any
nonzero */job_setup waiter, a following test's setup can pass/race against the dead
pulsar's registration.

Fix

Deterministic relay teardown that mirrors the AMQP force-drop:

• Add _relay_setup_waiter_count() (returns None when the relay is unreachable, so
  "unknown" is distinguishable from a confirmed zero) and refactor
  _relay_has_pulsar_setup_waiter() to build on it.
• Add _wait_relay_setup_waiters_drained() — polls /messages/poll/stats until zero
  */job_setup waiters; best-effort and bounded so it can never hang the suite.
• stop() and kill() await this drain in relay mode, so the following
  wait_until_consuming() only ever sees the freshly-started pulsar.

Harness-only change; no production code touched.

Testing

• py_compile and flake8 (max-line 150, complexity 14) pass.
• Not run against the live docker-compose stack — the failure is intermittent, so a
  single green run wouldn't prove the fix.

An independent interim mitigation (bind-timeout bump + fixture-setup retry) is proposed
separately in the 469-resilience-timeout PR; the two are non-overlapping and can land
in either order.

🤖 Generated with Claude Code

[mvdbeek]

Green on https://github.com/mvdbeek/pulsar/actions/runs/28662443620/job/85006053383, CI is severly hammered so I'll merge.

[ksuderman]

CI is severly hammered so I'll merge.

That would be me opening a flurry of pull requests all at the same time 😁. I am done now though and I'll stage bulk changes more slowly in the future.