The "race" to be the first thread using a socket is caused by the slave
views inheriting "notify yes;" from global configuration.  It causes a
NOTIFY to be immediately sent back to view "a" from views "b" and "c"
after view "a" notifies views "b" and "c" about a zone update.  As these
duplicate NOTIFY messages serve no purpose in this check, disabling them
will stabilize the test by working around the kernel bug mentioned.

While investigating this issue, I also noticed that junk gets logged
when using "also-notify { <ip> key <keyname>; };".  This is caused by a
misplaced call to isc_sockaddr_format().

Please review rt46516.  It fixes both problems mentioned above.  IMHO,
these fixes should be merged as two separate commits without a CHANGES
note, but if anyone feels otherwise, I have no strong feelings about it,
so please just let me know.