Huber/Robust regression; optimize sigma

[tlienart]

Huber regression in its default format works fine; Sklearn optimises an additional scale factor;

Probably can get to something similar by doing r ./= sigma and having a gradient in sigma etc; needs a bit of thinking if only because sklearn ensures sigma is positive by using LBFGSB and not just LBFGS.

Probably something that could be added optionally leaving the possibility for the user to CV it. Needs a bit of proper thinking as to blocking delta, blocking delta * sigma etc.

Allegedly it might make more sense to keep things as they are and CV the delta (or at least keep this possible) Though not directly comparable to sklearn as a result.