Ceph

This is necessary because though with ReplicatedBackend, log_operation will be called from within the submit_transaction call, with ECBackend, if a read is necessary to get a partial chunk, we will delay calling log_operation until we submit the transaction locally, possibly several message exchanges in the future. Thus, ReplicatedPG must track a pending_last_update which it looses when peering restarts. This should be used to initialize OpContexts.