Ensure k/v requests to memcached buckets use data nodes

"The basic issue is that the ketama hashing algorithm considered “all nodes” to be data nodes, and as a result would simply employ the node count to make use of this.
When we switched over from “nodes” to “nodesExt”, non-data nodes were placed into nodesExt starting in sherlock.

This becomes an issue when doing ketama hashing, in the fact that the hashing (if used against “all nodes”) will yield a mapping to a node that isn’t really a data node. Of course the solution is to only consider the count of the data nodes when performing this mechanism, but this is probably something we should fix and release note, and otherwise clarify.
"