Make initial draws more robust

Currently, when getting initial draws from the prior we check that the points have a finite likelihood/prior.

If the likelihood calls np.nan_to_num before returning (which I remember being necessary for some samplers) these points won't be caught.

I suggest catching these