Improve config refresh selection strategy

Right now the logic on how to select the refresher is very simplistic: if we loaded the original config over http, it will use http for refresh. If it loads through carrier it will use carrier for refresh.

Unfortunately, this is too simplistic: we might bootstrap off of a node in a MDS environment where there is no KV service enabled, but throughout the cluster there are. If it is a couchbase bucket and a modern cluster, it is more or less a given. Using http refresh has a couple limitations and should not be used if we can help it.