This patch splits the single operation of replace peer into two distinct
phases, in order to make a single change to the instance configuration:
  - detach from network
  - if at least one drbd device succeeded, make one change to the
    instance, pointing all disks to the new peer
  - attach to the new peer

This should eliminate the case when a failure in the second reattach can
leave the instance with two secondaries.

Reviewed-by: imsnah