Fix thread safety issues in beekeeper time_manager

- Change stop_requested from bool to std::atomic<bool> to fix data race
- Use condition variable instead of sleep_for to enable immediate thread
  wakeup during destruction
- Properly synchronize thread shutdown to prevent race conditions during
  cleanup

This fixes flaky beekeeper_timeout_unlock test failures caused by the
thread potentially accessing destroyed objects during shutdown.

Cherry-picked from hive commit 51d26ea24