For larger graphs (>= 50 nodes), reduce the accuracy and set a
minimum push flow threshold.
This avoids wasting a lot of time calculating very small flow demand
increments when the total number of demands is high and most demands
are small already.
Use a btree_map instead of a SmallMatrix.
This avoids excessive storage requirements for large graphs,
and overly expensive edge matrix resize operations.
Remove next_edge iteration mechanism.
