memory spike due to growing number of replica checkpoints if some of the nodes have slow disk write

a solution proposed for this issue is to coalesce checkpoints in the replica side if number of open checkpoints is more than two

https://github.com/membase/ep-engine/commit/dd4d60b79008586247e86c62e9ad2cc9320867d9

fixed in branch-17 and is going to be given to users as hotfix as need basis