Fix many NaNs in conditional parameter-spaces

[shukon]

When using conditionals, there might be nan-values... rendering the importance of an entire parameter useless. On the other hand, this solution might consider some parameters more than others by reducing the considered values? I am not sure how to handle this, @AndreBiedenkapp

[AndreBiedenkapp]

I thought the way the neighbors are sampled only allows for valid configurations in LPI.
Has something changed in the config space or am I simply mistaken?

But to give you an idea of how conditionals are handled in the other methods. In ablation we handle conditionals by jointly flipping them with their parent parameter. And then we attribute the importance to that tuple and not one individual parameter.
AFAIK fANOVA ignores conditionals entirely and pretends they are just normal parameters.

[shukon]

Well, the NaN's might have a different reason. I will investigate further...