[EM] Manipulability correction for BTR-IRV

[Kristofer Munsterhjelm]

I implemented BTR-IRV relatively quickly because Robert mentioned it and 
I thought it would provide useful context. It turns out my quick 
implementation was a little too dirty and it eliminated losers 
incorrectly when there was a tie for last.

It used to choose one candidate tied for last vs one candidate tied for 
next to last for the bottom runoff. Now it picks two candidates tied for 
last instead, which, in retrospect, seems much more sensible.

It improves the stats too. As with IRV, there's a candidate-dependent 
tiebreak but it shouldn't matter because the generator is candidate 
agnostic. Here's the new result:

Burial, no compromise:  62219   0.124438
Compromise, no burial:  2340    0.004680
Burial and compromise:  157     0.000314
Two-sided:              102669  0.205338
Other coalition strats: 1683    0.003366
========================================
Manipulable elections:  169068  0.338136

Ties: 0
Decisive elections: 500000

Condorcet//Plurality has a manipulability value close to ordinary 
Plurality, but I think my tie handling is obscuring the improvement from 
the Condorcet phase. So I should find a way to deal with ties before I 
post its detailed stats.

So much to do and so little time.

What I *can* conclude is that BTR-IRV is not the same thing as 
Condorcet-Plurality with four candidates or more. I kinda knew that 
already though :-)

-km