Cohen's Kappa

Cohen's kappa measures agreement between two raters beyond what chance would give:

where $p_o$ is observed agreement and $p_e = \sum_k p_{a,k}\, p_{b,k}$ uses each rater's label frequencies.

Implement cohens_kappa(a, b) for any label set. If $p_e = 1$ (both raters always say the same single label), return 1.0.