Poor scheduling throughput when a low percentage of tasks are schedule-able.

### Apache Airflow version 2.6.0 ### What happened Airflow scheduling task throughput suffers with relatively moderate numbers of total outstanding tasks - in situations where you have a few thousand tasks associated with running dags but a low fraction of them are schedu-able the scheduler spends a large portion of the time considering tasks before it finds the one or two that are eligible for scheduling in every cycle of the scheduler loop. Similarly, a dag with many blocked tasks can adversely impact a generally simple dagbag with moderate throughput requirements. In practice, the scheduler under nearly ideal circumstances can consider roughly 500-1000 tasks a second. This appears to be a combination of two factors. 1. Airflow is considering all tasks in that belong to a dag and re-considers each task each run of the scheduler instead of in response to events which can cause state changes. 2. Airflow's per-task consideration (including ti_deps) in the main scheduler loop is 1-2ms per task from a world of all non-completed tasks and is not sufficient to brute-force through without impacting scheduler latency. ### What you think should happen instead Either: ti_deps should be stored in the database and modified in response to events. The main scheduler loop should pull only tasks where there are no unsatisfied task dependencies and non check task-dependencies at run time. OR: per-task consideration in the main scheduler loop critical path should be reduced extremely significantly. Do any of the ti_deps result in further database calls for individual instances? If not, this path may not be plausible without the scheduler taking out a broad-scoped lock and keeping substantial state in memory. ### How to reproduce Construct a dag bag with two dags: * one dag (dag_id=sleep-then-parallel) that sleeps for an hour and then spawns 5000 parallel trivial tasks * another dag(dag_id=serial-dag) that runs every minute and spawns a 20-deep series of simple tasks. Each step o…