Should not always clear all the failed replicas when getting a flush all request

As described in this section in the design doc:

https://docs.google.com/document/d/1WPSBwRqIPgTrPRM2wtzAe-3XUvA9S7auWy3Fy8xuZLc/edit#heading=h.8nfip05gosr

We need to make sure that all the edits which have been sent to the broken replica must have a less sequence id than the flush sequence number of the commit flush edit, then we are safe to recover it. Otherwise we may miss some edit.

The current implementation is not enough, we need to improve it.