Improvement to Dinic's algorithm code

[howsiwei]

Instead of performing both bfs (initialize dist) and dfs (find augmenting path) from source, we can bfs from sink but still dfs from source. Doing so avoids exploring useless nodes that do not take us closer to sink while finding augmenting paths and a crude benchmark shows 10-20% runtime reduction on random graphs. This also makes dist more consistent with the labeling function in Push-relabel algorithm.

[howsiwei]

Essentially https://www.cs.bgu.ac.il/~dinitz/Papers/Dinitz_alg.pdf Section 4

[stevenhalim]

Thanks @howsiwei , noted :O, I will have to study this improvement first and convert my baseline maxflow.cpp with this idea. Leave it pending first.