Deprecate AGD, replace it by damped Newton's method

[eugenioLR]

The AGD method calculates the regularization term using a radius that can be obtained from the gradients as described in the following paper:

https://pdfs.semanticscholar.org/891a/d4990f2b9362bbfd9086d3a94c1a8e85de43.pdf

[eugenioLR]

Deprecate this method as it is simply a damped Newton's method with a trust region method to calculate $\mu$, it doesn't deserve to be a separate algorithm. Maybe Levenberg-Marquardt is the same, but perhaps it has enough recognition to deserve being a different method.