Instead of creating a new eventfd for each DeferredMonitor instance,
reuse EventLoop's eventfd, and add a std::list to EventLoop that
manages the list of pending DeferredMonitors.  This std::list is
protected by the same mutex as the "calls" list.

The bottom line is: reduced overhead because the per-instance eventfd
was eliminated, slightly added overhead due to Mutex usage (but
negligible), and we're thread-safe now.

This subsystem is now good enough to replace EventLoop::AddCall().