ZooKeeper should be managed by tests

[Gustavo Niemeyer]

Right now we're running tests against a hand-managed zookeeper. While this works, we're also constantly getting failed tests due to improper state inside zookeeper which was left around by a previously interrupted test. Rather than having to figure the problem by hand, and then resetting the state by hand too, the zookeeper server should be managed by a base class which is in charge of starting and stopping it, and resetting it whenever things go wrong.

The test zookeeper should run in a non-standard port, so that there's no danger of a test run ever killing a real deployment state.

This is being marked as high because we're constantly running into the interrupted test problem, and wasting time with it.