We already visit every single successor of the graph, so we don't

actually need to find the inverse edges, we can just store what they

are. This removes the need for inverse traits to compute dominators.

It's also a very slight speedup on large graphs, i'm leaving it here

while I test it on non-x86 platforms.