1D and 2D learner: stop interpolating function instead of the loss

Right now as the user requests more points but does not provide any data, we update the learner's state by adding "fake" points with values interpolated using a linear approximation. In the cquad implementation we realized that this is conceptually not right, and instead we directly consider the loss of the children intervals a fraction of the parent loss.

Also in the 1D and 2D learner this would make more sense and would allow them to support arbitrary loss functions.

Related: #50, !34 (closed).