Soft-sparsity localization of C^ug

[odunbar]

We should implement this: https://arxiv.org/abs/2208.03246

It has provable bounds on approximation with finite ensemble, and these extend to the case where it is imposed on the cross covariance too

[odunbar]

Tried this on orad/threshold-localizer by adding the new struct ThresholdCutoff(radius)

The outcome

it "can" work, but is incredibly sensitive to the radius. This is very likely due to the face that the cutoff is performed on the Covariance matrix and not the correlation matrix. We also find that even when applying it only to the correlation matrix, (equivalent to using our own SEC(0,radius)) localization, it still helps to have some additional damping, i.e. SEC(damp,radius).