Make weighting in pc conditional unbiased

[jhenderson0]

The current implementation uses group_weights2 / sum(group_weights2) for averaging over different conditioning groups. Instead we should used the same implementation as the probability of coincidence.