This covers the case of a multi-get targeting a downed server, getting pulled out and redistributed individually to a bunch of individual operations on other servers. Note that this is (should be): 1) Rare (you don't have servers dying a lot, do you? 2) Later optimized into per-server multi-gets when prepped for the wire.
KeyedOperations lets us know for which operations we know the original keys so we can do things like find another server to play the operations against.
This is *most* of the implementation of FailureMode.Retry.
This almost got profiling working in netbeans, but not completely.
The queue implementations were documented and should probably be reflected.
The external queue is still bounded, but the internal queues won't overflow by default (at the cost of more memory during spikes).
With this, observers can observe the *initial* connection, which (as the test demonstrates) can allow a caller to block until all (or at least enough) connections are established before proceeding.