The way we add or clear failedReplicas may have race

We will add a replica to failedReplicas in the callback of the replay method call, and we will clear the failedReplicas in add method when we meet a flush all edit.

They are in different threads, so it is possible that, we have already clear the failedReplicas due to a flush all edit, then in the callback of replay, we add a replica to the failedReplicas because of a failure of replicating, although the failure is actually before the flush all edit.

We should find a way to fix this.