duration of the whole 'flush instances'. As a consequence,
it was possible to query instance map, while it is in the
inconsistent state. The patch was reworked after review.


server-tools/instance-manager/guardian.cc:
  do not lock instance map in Guardian_thread::init()
server-tools/instance-manager/instance_map.cc:
  Eliminate race condition: lock instance map and guardian
  for the duration of the whole "FLUSH INSTANCES" execution.
server-tools/instance-manager/instance_map.h:
  add new method. cleanup interface. add comments.
server-tools/instance-manager/manager.cc:
  use instance_map.flush_instances instead of instance_map.load() and guardian_thread.init()