Re-organize read-loop execution to avoid crashes in read_loop on Win32 when ntp object destructs. Call quit on ntp_thread when ntp object destructs to free up thread (eventually we need to make fc::threads call quit in their destructor, but more work is required to make that work properly).
~fc::udp_socket now closes socket on destruction (consider doing this for tcp sockets as well).
- we were canceling tasks that hadn't been canceled, because the canceled flag was left set to true and the next task assigned to the context then became canceled as soon as it yielded
- we were resumeing blocked tasks before they should have resumed, because their blocking_promises list wasn't cleared and they were unblocking because the erroneous promises were fulfilled
As a debugging aid, we also record the cancellation reason whenever a task is canceled, and include that in the canceled_exception (this is only enabled in debug builds)
