The renderer benchmarks drain the main queue to quiescence before each
measured frame, so a task that yields and re-queues itself (an idle
sweep, a poller) runs to completion inside a single measured frame — or
keeps the benchmark spinning forever. Real platforms don't work that
way: macOS's main-queue drain only runs the items present when the drain
starts, Linux runs one runnable per calloop idle callback, and Windows
drains under a 10ms budget before servicing paint messages — frames
always get a chance to preempt. So `run_ready_main_tasks` now only runs
the tasks that were queued when servicing began, letting self-re-queuing
work advance one batch per frame like it would under a real run loop,
and `BenchAppContext` grows a `run_until` (wrapping the dispatcher
primitive task benchmarks already use) so setup code can pump the app
until a condition holds without settling to idle first.

Release Notes:

- N/A