Clean up some bugs in the callback handling

[resuna]

In some rare cases if you open and close a connection fast enough, synchronously, the close can happen before the callback gets through the zookeeper event queue and the Tcl event queue. This fix handles both ends of the problem:

1. Don't set up the callback if you're performing the init synchronously. The previous code put an event on the queue and exited the event handler early if there was no callback routine. This passes null as the event handler for zookeeper_init if there's no callback.

2. Don't zap the magic number in the zookeeper object until after calling zookeeper_close, because its cleanup may fire any pending handlers.