Finally fixed the bug where marathon were not releasing the lock

This was appearing because the KillStreamWatcher was waiting for all
tasks to die, and not only the overdue ones.

This was appearing most of the time when tasks were Unreachable because
they sometimes re-appear while marathon already scheduled new ones.

JIRA: MESOS-4345

src/main/scala/mesosphere/marathon/MarathonSchedulerActor.scala

@@ -452,7 +452,7 @@ class SchedulerActions(
 
     if (instancesToDecommission.nonEmpty) {
       logger.info(s"Adjusting goals for instances ${instancesToDecommission.map(_.instanceId)} (${GoalChangeReason.OverCapacity})")
-      val instancesAreTerminal = KillStreamWatcher.watchForKilledTasks(instanceTracker.instanceUpdates, instances).runWith(Sink.ignore)
+      val instancesAreTerminal = KillStreamWatcher.watchForKilledTasks(instanceTracker.instanceUpdates, instancesToDecommission).runWith(Sink.ignore)
 
       // Race condition with line 421. The instances we loaded might not exist anymore, e.g. the agent
       // might have been removed and the instance expunged.