- Fix Style/RescueModifier in poller_spec.rb
- Exclude oauth_token_refresh integration test from CI (requires
  separate Keycloak + OAuth Kafka infrastructure)

- Add separate CI job for OAuth integration testing with Keycloak +
  Kafka SASL/OAUTHBEARER infrastructure
- Lower minimum coverage to 95% since thread mode and FD polling mode
  exercise different code paths
- OAuth test runs in both thread and fiber polling modes

The Confluent Kafka image requires KAFKA_OPTS to be set when SASL is
enabled. This adds the required JAAS configuration file and mounts it
into the Kafka container.

To diagnose the OAuth token refresh failures in CI.

The listener-specific settings weren't being applied correctly.
Using global SASL OAUTHBEARER settings instead.

1. Fix token lifetime_ms to be absolute epoch time, not duration
2. Allow HTTP URLs for JWKS endpoint (Kafka 3.0+ security feature)

The allowed URL must match the full JWKS endpoint URL exactly.

Kafka needs Keycloak to be fully ready before it can fetch JWKS.
Using healthcheck with condition: service_healthy.

Keycloak takes longer to start on CI. Increased retries to 60
and start_period to 60s.

The Keycloak container may not have curl installed.
Using bash's built-in /dev/tcp for HTTP checks.

Tests that statistics events write to the queue FD and wake up
the poller, rather than relying on the 1s IO.select timeout.

Uses 100ms statistics interval and verifies we get ~30 callbacks
in 3 seconds (not just ~3 from timeout alone).

State already stores the monitor reference. Added attr_reader :monitor
to State and access monitors through states instead of separate cache.

Also adds integration test for statistics FD wakeup.

When the last producer is unregistered, set @shutdown = true to
stop the polling thread immediately. This prevents resource leakage
and aligns with how users are instructed to close producers.

The thread will restart automatically when a new producer registers.

Tests that:
1. Thread starts when first producer registers
2. Thread continues with multiple producers
3. Thread continues when one producer closes (others remain)
4. Thread stops when last producer closes (no leakage)
5. New producer works after all were closed (restart scenario)

- Fix FD_POLLING env check to use == "true" instead of truthy check
- Add missing polling.fd.periodic_poll_interval_format error message
- Remove unused poller.started/stopped events from notifications
- Add producer_id to broadcast_error for consistent error payloads
- Fix potential deadlock when unregister called from poller thread
- Update outdated comment about singleton instantiation in producer.rb
- Fix CHANGELOG benchmark numbers (44-64% -> 39-54%)

…e instance_variable usage from tests

macOS on GitHub Actions has less precise timing than Linux.
Increased sleep durations and timing thresholds to prevent
flaky test failures:
- connection_pool_spec: 0.01s -> 0.05s for thread sync
- idle_disconnector_listener_spec: 0.1s -> 0.2s for async disconnect
- state_spec: 20ms -> 100ms interval with proportionally adjusted sleeps
- producer_spec: statistics ts_d upper bound 200000 -> 300000

The 0.05s sleep wasn't giving thread1 enough time to grab
the connection on macOS before the main thread tried to
acquire it, causing the timeout test to fail.

Replace flaky sleep-based synchronization with a Queue to ensure
thread1 has actually grabbed the connection before the main thread
tries to acquire it. This eliminates the race condition that was
causing failures on macOS.