3.x: Implement lazy replica shuffling

[dkropachev]

It is followup for: #427
According to scylladb/cassandra-stress#38 (comment) there is still difference between Shuffled and not shuffled replicas, which made me think that array shuffling operations are very expensive.

Solution to that would be lazy shuffling, when original replica list is preserved as is while next item is being randomizes only when it is needed.
We need to do that and see if there is performance difference.