We can ignore when ConditionalDistribution.fit(...) fails for gridpoints. However, this currently returns the first guess - so in the result it seems that these points are valid as well. Should consider returning nan for them (but no idea if the rest of the machinery can handle nan).