isc_socket_sendto() calls that did not yet have their
completion events posted, so it could (rarely) end up
destroying the query context and then want to use
it again when the send event posted, triggering an
assertion as it tried to cancel an already-canceled
query. (RT #77)
... so now the resolver counts the number of outstanding isc_socket_sendto()s
and only destroys the query context when there are no more.