Improve our strategy for preventing recovery requests from stacking up.
Zookeeper crash JVM stack OOM because of recover strategy
Stacked recovery requests do no cancel an in progress recovery first.
Add a new test fault injection approach and a new SolrCloud test that stops and starts the cluster while indexing data and with random faults.
RecoveryStrategy should not publish any state when closed / cancelled.
Harden the 'leader requests replica to recover' code path.