Improve modeling of covariance matrix

[anicola]

Improve the modeling of the covariance matrix, especially by including errors due to uncertainty in theory predictions, and possibly non-Gaussian terms.