Speeding up partition reassignment after broker failure

After a broker failure the controller does several writes to Zookeeper for each partition on the failed broker. Writes are done one at a time, in closed loop, which is slow especially under high latency networks. Zookeeper has support for batching operations (the "multi" API). It is expected that substituting serial writes with batched ones should reduce failure handling time by an order of magnitude.

This is identified as an issue in https://cwiki.apache.org/confluence/display/KAFKA/kafka+Detailed+Replication+Design+V3 (section End-to-end latency during a broker failure)