Benchmark questions related to Preflow

[Peter Kovacs]

There are some efficiency questions about the Preflow implementation, that require thorough benchmarking.

• Elevator or LinkedElevator should be used by default?
• The BFS implementation in the init() function could be made more efficient using only one vector and three indices (first, last, new_level).
• We should try very large instances and check again whether the current heuristics and the constant 20 factor that is used for them are really optimal or not.
• We should try Goldberg's new idea of partial augmentations. (However it should probably be a new class.)

[Peter Kovacs]

Another idea to be considered is to store the last outgoing arc for each node, see #431.