As we don't (and must not, otherwise it would deadlock) take a write lock
on a storage node when detecting a conflict, it is possible that multiple
conflicts get successively reported for a single object in a single
transaction. To solve this, client must tolerate multiple conflict for a
single object. The last resolved conflict being sent to all storage nodes,
transaction data will be consistent at tpc_finish.

- conflict_serial_dict and resolved_conflict_serial_dict become
  multivalued (sets instead of scalar)
- each resolution only resolve conflict with highest serial if greater
  than previous resolutions for the same object
  NOTE: this assumes that ("r" being conflict resolution method, "Tn"
  being the TID on which object is based, and "Cn" being a conflict with
  transaction n) when r(T0, C1) generates T1 but triggers C2,
    r(T0, C1) + r(T1, C2) == r(T0, C2)

git-svn-id: https://svn.erp5.org/repos/neo/trunk@2108 71dcc9de-d417-0410-9af5-da40c76e7ee4