Update covariance

[wdevazelhes]

In the Covariance algorithm, we currently do an inverse of the covariance matrix, and the latter could be singular. We need to decide whether to return an error or return the pseudo inverse in this case.

[bellet]

I think we should use the pseudo-inverse, which is a common thing to do, see for instance the following link (sorry for the lack of better reference):
https://stats.stackexchange.com/questions/37743/singular-covariance-matrix-in-mahalanobis-distance-in-matlab