sanlock: correct paxos usage
    
    Rework how the disk paxos algorithm is applied to the
    ownership and updating of the leader_record.  How this
    was done and worked previously had never been thoroughly
    understood and was wrong in a number of general ways.
    Some main concepts driving these changes:
    
    . ballots are run to commit each new lver
    . a node that successfully completes the ballot needs to
      commit the result from that ballot, which means
      writing the lver from that ballot and the dblock values
      to the leader_record
    . nodes need to monitor the leader_record to learn if
      the next lver has been commited, and if so, return that
      result
    . the hosts participating in a ballot are the ones with
      the latest matching lver in their dblocks
    . each node's mbal should always increase
    
    This includes a new incompatible disk format change,
    due to the addition of some extra info fields to the leader
    and dblock structs.
    
    Error codes are also changed to describe the new conditions.