Use sacc in multi-like covariance

[mgerbino]

We should upgrade the reading of the cross-covariance in the multi-like module so to use sacc format.

[itrharrison]

Note: ensure sacc is used where possible, but do not require sacc use where currently not possible (e.g. cluster counts, lensing noise).

[mgerbino]

@itrharrison Currently, cross-covariances (when available) are loaded from a .npz file that contains several issues of cross-cov. Since sacc allows to store one covariance per file, maybe sacc is not the most suitable format in this case? We can keep npz file or replace it with a fits.

[itrharrison]

I think sacc can select only relevant blocks of a covariance matrix... so we would store the whole covariance matrix and then just extract the right bits if not using all the probes.

So if you remove some tracers or some data points then I think the parts of the full covmat are also removed along with the data points.