bug: make sampler work when NaN values are predicted by the models

[younesStrittmatter]

Sampler was breaking with an Error, if one of the models predicts NaN or Inf for a single condition.

Fix:
The disagrement (distance) between two models get's set to 0 if at least one of them predict NaN or Inf at that condition.

Rationale:

• Throwing an Error is no good, since it breaks with a lot of models
• Dropping the condition is no good, since some models would "clog" the sampler (For example, if a model is sqrt(x), this means no negative conditions get sampled anymore.
• Setting the distance to a high number would lead to overly sampling potentially uninformative conditions (especially if the ground truth actually is undefined at these condition)
  -> Setting the distance to zero doesn't encourage sampling this condition, but also doesn't make it impossible.

[musslick]

Looks great. Only consideration: Do we want to issue a warning once a nan is detected, so that the user is informed about how the disagreement is computed in this case?

[musslick]

Looks great!