Complex fixes

[kagalenko-m-b]

The proposed changes aim to repair a few issues with levenberg_marquardt():

• At present, it allows complex nonlinear variables. However, squared norm is not analytic, and optimizing it requires taking partial derivatives with respect to both z and conj(z). Best not to confuse users and simply disallow complex variables, recommending instead to treat them as pairs of real and imaginary parts.

• It does not handle complex-valued Jacobian correctly

• Computation of LM step involves forming normal equations, which squares the condition number.

[pkofod]

Thanks, I will try to look at this but I have been very busy lately unfortunately.

[pkofod]

This week is quite busy, but I promise to get back to you. Maybe @antoine-levitt has some comments in the mean time?

[antoine-levitt]

I'm not familiar with the algorithm so I just left trivial style comments. Generally treating the complex optimization problem as real/imag is a good idea (sometimes it's unnecessary, but it's always a good fallback choice)