Update datanode replacement policy to make writes more robust

As discussed in HDFS-5924, writers that are down to only one node due to node failures can suffer if a DN does not restart in time. We do not worry about writes that began with single replica.