Improve MCMF

[Matistjati]

This PR improves 3 things:

1. Fixes the time complexity description, as mentioned in Wrong Time Complexity for MinCostMaxFlow #217
2. No longer uses quadratic memory (adjacency list instead of matrix). Quadratic memory was problematic on problems with sparse graphs such as tourist.
3. Allows parallell/double edges.

Empirically, it also seems to improve performance (see "improved MCMF" here)

Potential downsides of new implementation:

1. Loops (edges from a node to itself) don't work (when would you ever want that, though?)
2. Uses pointers to easily reconstruct path, making some parts less readable.

It passes all stress tests and doesn't get any WA's on the kattis problems I've tried. It might be worth to consider using SPFA for setpi if it does not complicate the code too much.

Side note: does anyone know why we can skip using setpi if input costs are positive?

[simonlindholm]

Nice to hear this improves perf, I was worried about that before but I guess it makes sense. (Maybe it would another thing if the vector<vector<>> was a packed matrix.)

I agree self loops aren't interesting -- they could be if they had negative weight but we don't handle negative weight cycles anyway. And them not showing up in the list of edges doesn't feel like a problem at all since you'll be filtering out the subset of edges with flow > 0 anyway if you ever go through it.

[simonlindholm]

If you make fast to create kactl.pdf you see this line overflowing from having > 63 chars; any way it could be shortened to keep it on a single line? E.g. shorten par or from or trim some whitespace or move the declaration of p out. Or actually, remove p since it doesn't seem to be used other than in one place? There are some other lines that just barely overflow too.

[Matistjati]

I chose to use python-style ifs to fix the "push into priority queue" from overflowing. A bit stylistically inconsistent, but probably fine?

kactl/content/graph/MinCostMaxFlow.h

Lines 53 to 56 in 48ca736

	if (its[e.to] == q.end())
	its[e.to] = q.push({ -dist[e.to], e.to });
	else
	q.modify(its[e.to], { -dist[e.to], e.to });

After fiddling with the loops, I found a pretty clean solution in the end. All overflows should be fixed now.

kactl/content/graph/MinCostMaxFlow.h

Lines 67 to 74 in 48ca736

	for (edge* x = par[t]; x; x = par[x->from])
	fl = min(fl, x->cap - x->flow);
	totflow += fl;
	for (edge* x = par[t]; x; x = par[x->from]) {
	x->flow += fl;
	ed[x->to][x->rev].flow -= fl;
	}