Uses strong connection information for specifying edge weights when

repartitioning.
Do not skip isolated nodes. (In parallel they might occur on coarser
level due a not ideal data decomposition)

[[Imported from SVN: r1080]]