Standardize implementation of log-likelihood

We should use Ted's formulation of log-likelihood instead of the one that exists twice in the CF module. It doesn't change the resulting numbers, though does fix some symmetry issues in corner cases.