fix use-after-free of the timer node mutex on destruction

[aysha-afrah26]

The Sleep, Delay, and Timeout nodes each own a TimerQueue whose worker thread runs the timer callback, and that callback locks a std::mutex that belongs to the node (delay_mutex_ or timeout_mutex_). Because the TimerQueue was the first-declared member it is destroyed last, after that mutex, so when a tree is torn down while a timer is still pending the worker thread (joined only in ~TimerQueue) can lock a mutex whose destructor has already run. On macOS/libc++ the pending handler then aborts the process with "mutex lock failed: Invalid argument"; on other standard libraries it is undefined behavior. It shows up as intermittent aborts in the timer and parallel teardown tests, including Parallel.Issue819_SequenceVsReactiveSequence, which the macOS CI currently skips as flaky. Declaring the TimerQueue as the last member fixes it, since ~TimerQueue() joins the worker thread while the mutex it locks is still alive. Runtime behavior for valid trees is unchanged; only the member declaration order, and therefore the destruction order, moves.