Allow y_hat instead of unfitted model as input for regression ICP

[ghost]

Users may have predictions generated from different fitted underlying models in different languages and just want to get the prediction interval. This change allows user to set underlying model = None but using y_hat (prediction value) to get the result. Only regression part is done in this commit for my work purpose. Will enhance classification part as well later.

[donlnz]

I agree that it seems reasonable to provide functionality to produce prediction intervals from a model trained outside of nonconformist, but I'm not sure this is the 'nicest' way. Ideally, in situations like this, I would recommend creating a custom ModelAdapter (inheriting from, e.g., nonconformist.base.RegressorAdapter). Would that be a possible alternative in your use case?