Revision 1023181

HBASE-3114 Test up on hudson are leaking zookeeper ensemble

Trying this patch to see if it cures hudson of leakage

M src/test/java/org/apache/hadoop/hbase/HBaseTestingUtility.java
  Add a passedZkCluster flag.  If we didn't start the zk cluster,
  don't shut it down on way out.
M src/test/java/org/apache/hadoop/hbase/TestHBaseTestingUtility.java
  Add a test that starts up three clusters in a vm all sharing single
  zk, each to its own chroot location.  Add to tables in each and
  very the add by doing same as the old TestMultiClusters used)
M src/test/java/org/apache/hadoop/hbase/replication/TestReplication.java
  Reenable shutdown of started clusters.  Hopefully this clears
  the leaking of zk ensembles up on hudson.
M src/main/java/org/apache/hadoop/hbase/zookeeper/ZooKeeperWatcher.java
  Testing, it looks like legitimately the callback for session connected
  can be invoked before construction of zookeeperwatcher finishes.  In
  particular the zookeeper data member can be null.  Hang around for a
  second of so before throwing an exception (Make the exception
  indicate the particular zkw by adding to the error message stack
  trace made at zkw construction -- helps with debugging this stuff)