Use single vector of edges rather than per-node vector. There is a small
hit to instruction counts (< 0.5%), but the memory savings make up for it.