There were three surprising things in the previous version:
 * the XXX mention: leafs can indeed be detected at construction time
 * the need to track both children and parents of a node: usually only one
    or the other is enough to work efficiently on a graph
 * the idiom:
      if n in list: list.remove(n)
      list.append(n)

Replace it by a generic sort on node depth. A dictionary is used for lookups to
avoid linear list lookups, and only the successor relation is kept (parent_dict)

Also explicit in comments the graph we are building/the order expected for code
clarity (ie avoid confusion wrt what is the parent/child of the graph node)


git-svn-id: https://svn.erp5.org/repos/public/erp5/trunk@31556 20353a03-c40f-0410-a6d1-a30d3c3de9de