Substantially based on a patch by Matthias Clasen <mclasen@redhat.com>
kqueue implementation by Joe Marcus Clarke <marcus@freebsd.org>

Previously, when we detected a configuration change (which included
the set of config directories to monitor for changes), we would
simply drop all watches, then readd them.

The problem with this is that it introduced a race condition where
we might not be watching one of the config directories for changes.

Rather than dropping and readding, change the OS-dependent monitoring
API to simply take a new set of directories to monitor.  Implicit
in this is that the OS-specific layer needs to keep track of the
previously monitored set.