make failure injection idempotent

- instead of creating local top-level actors per pipeline, just create
  one system actor through which everything is sent
- this enables storing settings (like what to throttle how) within this
  actor and applying settings when connections come up later
- it also gets rid of the blocking actor creation from
  NetworkFailureInjector, fixing the dead-lock
- moved also the ServerFSMs to be children of the Controller
- all actors have proper names now for easier debugging

akka-remote/src/main/scala/akka/remote/netty/Client.scala

@@ -173,6 +173,7 @@ class ActiveRemoteClient private[akka] (
         notifyListeners(RemoteClientError(connection.getCause, netty, remoteAddress))
         false
       } else {
+        ChannelAddress.set(connection.getChannel, Some(remoteAddress))
         sendSecureCookie(connection)
         notifyListeners(RemoteClientStarted(netty, remoteAddress))
         true
@@ -196,8 +197,10 @@ class ActiveRemoteClient private[akka] (
 
     notifyListeners(RemoteClientShutdown(netty, remoteAddress))
     try {
-      if ((connection ne null) && (connection.getChannel ne null))
+      if ((connection ne null) && (connection.getChannel ne null)) {
+        ChannelAddress.remove(connection.getChannel)
         connection.getChannel.close()
+      }
     } finally {
       try {
         if (openChannels ne null) openChannels.close.awaitUninterruptibly()