MM-25154: Fix data race in InvokeClusterLeaderChangedListener (#14610)

The order of booting up a server should be
Job scheduler -> Cluster.

And shutdown should be the opposite. This is needed because the job scheduler
initializes certain data structures that are later called by ClusterLeaderChanged
event handlers. And the event handlers run in a separate goroutine.

Therefore, if the cluster initialization happens before the job scheduler,
then a race happens between accessing the jobs variable and setting it.

To prevent this race, we fix the order of startup, and add comments in both
places to prevent things from regressing again.

app/server.go

@@ -350,6 +350,9 @@ func NewServer(options ...Option) (*Server, error) {
 		mlog.Error("Error to reset the server status.", mlog.Err(err))
 	}
 
+	// Scheduler must be started before cluster.
+	s.initJobs()
+
 	if s.joinCluster && s.Cluster != nil {
 		s.FakeApp().registerAllClusterMessageHandlers()
 		s.Cluster.StartInterNodeCommunication()
@@ -374,8 +377,6 @@ func NewServer(options ...Option) (*Server, error) {
 		}
 	}
 
-	s.initJobs()
-
 	if s.runjobs {
 		s.Go(func() {
 			runSecurityJob(s)
@@ -490,6 +491,7 @@ func (s *Server) Shutdown() error {
 		s.Metrics.StopServer()
 	}
 
+	// This must be done after the cluster is stopped.
 	if s.Jobs != nil && s.runjobs {
 		s.Jobs.StopWorkers()
 		s.Jobs.StopSchedulers()