This changes the underlying adjacency matrix implementation to a single
mixed in + out edge list for each node, instead of separate ones. This
simplifies the undirected_adaptor<> so that its out-edge iterators
become trivial, and also faster.