Implement graph-based conservative check for counterfactual fairness

[zenna]

As mentioned in the paper, a sufficient condition for a model being counterfactually fair is that the causal descendants of the protected attributes are not used in the classifier.

Implement a graph based check for this