The problem with previous implementation was that each storage locally
decided what undo actually did to data. This causes problems when a
storage doesn't have a complete view of past transaction but accepts write
queries, ie when it replicates.
This implementation reduces the decision to a readable subset of storage
nodes (which are hence not replicating), and then sends that decision to
all storage nodes, hence fixing the issue.
Also, DatabaseManager.storeTransaction now consistently expects object's
value_serial to be packed, not an integer.

git-svn-id: https://svn.erp5.org/repos/neo/trunk@2285 71dcc9de-d417-0410-9af5-da40c76e7ee4