Singleinflight interferes with the health checking of upstream. If an
upstream would fail, singleinflight would mirror that error to to other
proxy *iff* multple identical queries would be inflight. This would lead
to marking *all* upstreams as bad, essentially collapsing multiple
upstreams into a SPOF. Clearly not what we want.

Singleinflight does have some nice properties, but I've opted to rip it
out entirely. Caching should almost (but not quite) as good.

Added a test case in test that uses 3 CoreDNS instances to reflect the
setup from #715.

Found another bug as well, where (when the policy would be nil), we
would always Spray even though we've found a healthy host.