Fix stuck faststream-stomp consumers (ping=True but no progress)

[lesnik512]

Summary

In production, faststream-stomp consumers stop processing messages while broker.ping() keeps returning True. Queue depth grows on the broker (ActiveMQ Artemis), the DLQ also grows, and only a pod restart recovers.

This PR addresses every failure mode that can lead to that state:

• Listener safety net — wrap every per-message handler so a single handler crash can no longer cancel the read loop. Previously, an unhandled exception inside a handler would cancel the inner TaskGroup and silently kill the listen task.
• is_alive grace fix — ActiveConnectionState now tracks connected_at; when last_read_time is None, aliveness is bounded by the same server_heartbeat × factor threshold instead of returning True indefinitely. Closes the indefinite-True window after a reconnect where nothing reads.
• Listen-task death surfaces in is_alive — Client.is_alive() returns False once _listen_task.done(), so dead listeners can no longer report healthy.
• Unhandled listen-task exceptions are logged — _handle_listen_task_done now logs every non-cancelled, non-FailedAllConnectAttemptsError exception at ERROR with the traceback. Previously these were silently swallowed.
• max_concurrent_handlers knob — opt-in semaphore (default None/unbounded) bounds in-flight handler tasks. When the semaphore is full the read loop pauses, which (combined with the is_alive fix above) converts an invisible stuck-state into a detectable one.
• Visibility for ack/nack drops — ack-skip-after-reconnect → WARNING; nack-skip-after-reconnect → ERROR. Dropped frames inside maybe_write_frame are also logged with severity by frame type (NACK=ERROR, ACK=WARNING, other=INFO) so DLQ growth becomes diagnosable from logs alone.

Test plan

• just lint
• just check-types
• just test-fast — 244 passed / 1 skipped
• just test — full suite (requires Docker brokers); run before merge

Notes for reviewers

• Default behavior is preserved: max_concurrent_handlers is None by default; existing call sites do not need changes.

[vrslev]

Perhaps, we should set a sensible default for max_concurrent_handlers?