There are 2 objectives:
- Prevent randomly trying to connect to an unresponsive storage node, which
  impairs performances a lot. Note that this happens only when the master
  didn't notice the disconnection, so the node is still in running state in
  the node manager.
- Increase connection reuse, saving the cost of establishing a new
  connection and a slot in connection pool.
Randomisation should be kept to even out storage node use.

git-svn-id: https://svn.erp5.org/repos/neo/trunk@2173 71dcc9de-d417-0410-9af5-da40c76e7ee4