[server] fix zk notification purge race

[gyang94]

Purpose

Linked issue: close #3753

Brief change log

empty check for zk stats

Tests

API and Format

Documentation

[fresh-borzoni]

@gyang94 Thank you, fix is sound, some ideas about the test below, PTAL

The test passes even with the fix reverted because our test config has the root logger OFF, so the appender never gets the ERROR and noneMatch on an empty list is always true.
So a couple of suggestions:

1. mockito isn't really needed: the handler runs before the purge and they share the children list, so a handler that deletes the notification node gives the purge a stale list, the exact race from the issue.
   a sketch:

 TestingNotificationHandler handler =
         new TestingNotificationHandler() {
             @Override
             public void processNotification(byte[] notification) {
                 super.processNotification(notification);
                 deleteAllNotificationNodes(); // purge now sees a stale list
             }
         };

2. with root OFF the appender only works through its own LoggerConfig, smth like this:

LoggerContext ctx = (LoggerContext) LogManager.getContext(false);
Configuration cfg = ctx.getConfiguration();
LoggerConfig lc = new LoggerConfig(ZkNodeChangeNotificationWatcher.class.getName(), Level.DEBUG, false);
lc.addAppender(appender, Level.DEBUG, null);
cfg.addLogger(lc.getName(), lc);
ctx.updateLoggers();

and assert the debug message is present rather than the error absent, so it can't go vacuous again.