Tasks remain in queued state for ~950s despite available worker capacity and no resource bottlenecks

### Apache Airflow version Other Airflow 2 version (please specify below) ### If "Other Airflow 2 version" selected, which one? 2.9.0 ### What happened? In our production environment using `CeleryExecutor` and Redis as the broker, we’re experiencing intermittent but repeatable delays where tasks remain in the `queued` state for approximately 950 seconds before they begin execution. This behavior is unexpected given that: - There is plenty of free capacity in the pools used by the tasks. - Four Celery workers are running, each with a concurrency of 32, for a total of 128 available slots. - The Redis broker is not under high CPU/memory pressure. - The are 4 schedulers (multi-scheduler enabled) and their heartbeat is healthy and operational. - We verified the tasks are coming from different DAGs and using different pools. - There is no CPU saturation on any component (scheduler, Redis, workers). - The logs from the scheduler and workers show no errors or timeouts. - The configuration key `AIRFLOW__SCHEDULER__TASK_QUEUED_TIMEOUT` is set to 3600. Despite this, some tasks get queued and then do not start executing until ~950 seconds later, even when there is no obvious bottleneck or backlog.  ### What you think should happen instead? We expected tasks to transition from `queued` to `running` within a few seconds, as there are ample resources and capacity available. ### How to reproduce We don't have a minimal reproducible example, but the issue appears randomly and independently of DAG size or the number of enqueued tasks. For example: 1. A DAG triggers one or more tasks. 2. Some tasks transition from `queued` to `running` quickly as expected. 3. Others remain stuck in `queued` for approximately 950 seconds, even when: - Pools have available slots. - Workers are idle or underutilized. - Redis is healthy and not under pressure. - There is no backlog of scheduled or queu…