Race condition in AbstractReplicationStrategy.getNaturalReplicas

There is a very narrow and infrequent race window, in which two ring updates occur in a short space of time (or during an interval of no queries):

• thread A invalidates the cache after the first ring change, snapshots this version of the ring, and begins to calculate its natural endpoints
• thread B sees the second ring change, and invalidates the cache before thread A completes
• thread A writes its value to the cache, based on the old ring layout

Now, a stale view of the endpoints for this token will be persisted in AbstractReplicationStrategy until the next ring change (which may feasibly never occur)