TODO

 X Coupling of parameters
 X Use other optimization algorithms
 X Check that initial parameters are inside the range if not inverted
 X Masks
 X Save chi2 for each cycle
 X Comment on virtual environments
 X Explain better how atmospheres are defined in Hazel
 X Randomization
 X Error bars
 X Allow to invert magnetic fields in the LOS or vertical reference frames
 X Instrumental profile
 X Add more tests
 X Output AIC and BIC (N-> ndata, Nvar -> n.variables)
   r_i = (data-model) / sigma
   chi2 = sum(r_i^2)
   aic = N*log(chi2/N) + 2*Nvar
   bic = N*log(chi2/N) + ln(N) * Nvar

 - Sparse regularization of maps : change order of loops -> 1 LM (x,y) -> regularization
 - Regularization of parameters
 - In synthesis, all atmospheres have to have the same number of pixels

 - Add utilities for generating new SIR models
 - Show an example for inverting a random distribution of points to be inverted and then
   start from an interpolation of these maps to be closer to the solution
 - Develop tools to easily manage input/output models so that multiinversions are easy to do.
 - Check that output and input models are compatible so that reinversions can be easily done
 - Precalculate coefficients for the emission/absorption coefficients using sparse
   matrices and use them later on the calculation. This will avoid many of the loops
   that are, by construction, zero.
 - Programmatic inversions