Writes lost during replica churn

[atamon]

Running #37 the following scenario may result in lost writes:

• All 3 replicas down
• replica 0 goes up, starts state transfer to 1 and 2. state transfer succeeds
• replica 0 and 1 goes down again
• clients write to replica 2
• Split brain, manual scale down required
• replica 0 goes up as "new cluster", begins state transfer to 1 and again
• writes written to replica 1 or 2 are lost