far.py: max-over-subset ln L

- redefine the ranking statistic to be the largest ln L for any subset of the triggers in a candidate to prevent a glitch in one detector from making a good candidate look bad