Fixes for shutdown sequence, including prevention of a race condition

* Throw an exception if any events, runnables or messages are posted after shutdown. These were previosuly being silently ignored. We really should know about them.
* Don't exit the connection manager thread until the object manager thread exits. This fixes a race condition where the wakeup time of the conmgr thread would determine if it shutdown before or after the object manager, meaning it would sometimes ignore messages around shutdown time.
* Add a shutdown constraint to make the client manager shutdown before the invoker/event/conmgr threads. This gives a single point in the shutdown sequence that applications can use as a constraint (TODO: shutdown groups?)
* Emit a log message for each shutdown method call. This could be very useful in production.

git-svn-id: svn+ssh://src.earth.threerings.net/narya/trunk@5491 542714f4-19e9-0310-aa3c-eee0fc999fb1

src/java/com/threerings/presents/server/ShutdownManager.java

@@ -127,6 +127,7 @@ public class ShutdownManager
         // shut down all shutdown participants
         downers.apply(new ObserverList.ObserverOp<Shutdowner>() {
             public boolean apply (Shutdowner downer) {
+                log.info("Calling shutdown registrant", "downer", downer);
                 downer.shutdown();
                 return true;
             }