For our flat lan case, we run lans directly over the top of real (or
emulab vlan) networks.  So we extend those emulab IP networks into
the openstack flat network associated with each emulab network.  Well,
openstack assumes it controls the subnet and can just allocate special
addresses anywhere -- like for the local dhcp agent IP, or a router
interface -- it doesn't respect the allocation_pool value we give the
network (so that must just be for compute nodes).

Further, there's no way to set these special IPs when creating networks,
subnets, and routers, so we retroactively find those ports and change
their IP addrs to something that won't stomp on emulab IPs, nor on the
openstack allocation_pool.

This seems to work great, except that packets outbound from an openstack
instance don't get SNAT'd so they appear they came from the real
external world address.  Inbound packets make it all the way, so at
least DNAT is working.