Investigate safer GlobalTask scheduling API

Related to the cause of MB-32181, currently the APIs to schedule, run and snooze GlobalTasks in ep-engine are brittle and easy to use incorrectly:

• GlobalTasks maintain their own wakeTime as a member variable; however these tasks are then referenced from a FutureQueue (std::priority_queue) in wakeTime order.
• If a GlobalTask has it's wakeTime changed; we must then re-sort the FutureQueue.
• If we don't re-sort (as was the case with MB-32181), then scheduling breaks horribly.
• By default a task will be re-scheduled immediately if the run() method completes (and the method doesn't return false to cancel the task).
• To override this behaviour the run() method must call snooze() or wakeUp() to modify the wakeTime - however those functions are public and can be called by anyone, even though it would be unsafe to call them if the FutureQueue isn't going to be re-sorted.

Possible improvements to this vary:

One modest change would be to change the run() method API instead of returning just a bool (true=schedule again, false=cancel), it returns an optional<time_point> :-

• If the time point is empty then cancel the task.
• If the time point is non-empty then it specifies the next time to schedule the task
  This would enforce that a task specifies exactly once, when it should be run again (and removes the immediate re-scheudule assumption which actually a lot of tasks don't want). The GlobalTask::wakeUp and GlobalTask::snooze methods could then be removed.

Another approach would be something more ambitious - maybe look to move to a totally new scheduling API such as folly::future or the wangle library...