This implements a retry policy to consistently retry another host within the SRV
set if we get a failure talking to the original host.

We still use the original fallback strategy if this retry fails. (Should we?)

This extracts the consistent hashing code into a new package and writes more
tests for it.

This removes a bunch of duplication from ConsistentHashingMode.

As a consequence this jumbles the consistent hashing algorithm (because I've
embedded the Key into a top-level CacheKey struct which includes Attempt for
retry support). At this stage this is safe because nothing live is using it.

rename method

avoid implicit return

remove unneeded conditional on m.DomainsToCache