Improve density cross-validation

[mdtanker]

The cross-validation procedure currently implemented for finding the optimal density contrast value is ineffective when there is any significant regional field. It always overestimates the density contrast. This is due to higher density contrast resulting in less change to the topography. When there is a regional field that is not fully accounted for, the higher density will lead to less incorporation of the regional signal into the inverted topography, and therefore a better CV score.

This seems to be mostly a problem with both the technique of using a non-constraint point minimization method of regional separation during the CV, then repeating the inversion with the best damping parameter and CPM.

This might be an inherent problem which is not solvable. Strangely, while the CV overestimated the density, it still chooses the density value which results in the optimal inverted topography.

[mdtanker]

This issue have been marginally improved in v0.10 from a bug found in optimization.optimize_inversion_zref_density_contrast_kfolds where the starting topography was not being recreated for each fold of constraint points.