fix: Fix flaky TaskQueueScaleTest.doMassLaunchAndExit

[Akshat-Jain]

Description

This PR fixes flaky TaskQueueScaleTest.doMassLaunchAndExit. Sample failure: https://github.com/apache/druid/actions/runs/36084362652/job/107912839489

   🧪 - indexing-service/target/test-classes/org/apache/druid/indexing/overlord/TaskQueueScaleTest.class | all tasks should be known ==> expected: <1000> but was: <973>
  Error: all tasks should be known ==> expected: <1000> but was: <973>

TaskQueueScaleTest.doMassLaunchAndExit reads the running, pending and waiting counts at three separate instants while the TaskQueue-Manager thread is still submitting tasks to the runner. Since getWaitingTaskCount() counts active tasks the runner does not know about, a task that crosses taskRunner.run() between the pending read and the waiting read is counted by neither, so the sum comes up short (expected: <1000> but was: <988>).

The test had a high failure rate without this PR change. With this PR change, I ran the test 100 times locally, it passes everytime.

---

This PR has:

• been self-reviewed.

[FrankChen021]

Thanks for the fix. However, a simple while-loop is not enough. It has been fixed by #20291
Let's see if we have similar flaky test after that.

BTW, under the GitHub issue section, there's a series issues created by Claude Code everyday that triages CI failures in previous day, some of those are still open, hope someone can fix them.