As I dont quite get the logic of "enforcing a positive quadrant" from
b38be90515 (which broke the previous behavior of 2.79 which was
capable of handling reversed edges as well -- there might be ways to
make that work though), now use the alternative approach of storing the
flipped direction in the kdtree as well.
This has a slight performance hit (only noticable at ~500k edges), but I
believe correctness beats performance here.