[EM] RP implemented on CIVS

[Andrew Myers]

I added a version of ranked pairs/winning votes (as I understand it) to the
CIVS voting system, though it's a little harder to get to than the beatpath
winner result. This is partly because RP seems to be much more expensive to
compute, even with some rather effective optimizations I added to incrementally
compute transitive closure.

The big quandary I faced was what to do with preferences of equal strength.  I
don't like random methods because I think they make voters less confident about
the results. So I adopted the following rule: a preference is considered as
part of the final preference graph if it does not create any _new_ cycles when
considered in combination with only _strictly stronger_ preferences.  In other
words, the final graph might contain cycles, but only cycles that arise because
preferences of equal strength create a cycle when they are accepted, whereas
each of those preferences individually would not create a cycle.

In various trials it seems to work well. I find that the result seems to be
more stable than beatpath winner in the sense that individual voters don't
perturb the output order as much. But this is not a result that I have
methodically tested, just a strong impression. I also find it easier to explain
and justify than beatpath winner. Anyone is welcome to try it out on the web
site, of course.

So I'm curious to know whether this is a variant of ranked pairs that has been
explored (or named) previously and whether there are any evident problems with
it.

-- Andrew