`ThreadedDispatcher::run_until` - the completion mechanism under
`BenchAppContext` async task benchmarks (`bench_task` /
`bench_batched_task`, added in #62180) - drained the entire main queue
before checking its readiness predicate. Main-thread work that re-queues
itself (an idle-time sweep, a polling loop) keeps the queue non-empty
until it finishes every iteration, so a task benchmark's measured
interval silently extended past the awaited task's completion until all
such deferred work settled. Benchmarks of UI workloads that schedule
follow-up idle work could report several times their true completion
latency.

This changes `run_until` to step main-thread runnables one at a time and
check readiness before each, so it returns at the completion it awaits
rather than at queue quiescence, making async task benchmarks more
accurate. The regression test spawns a task that re-queues itself 10,000
times plus a one-shot completion task, and asserts `run_until` returns
without draining the re-queued work; it fails against the previous
drain-first implementation.

Release Notes:

- N/A