Warn users who don't provide a real covariance matrix?

[lebigot]

No description provided.

[lebigot]

When defining numbers with uncertainty based on a covariance or correlation matrix, it could be useful to warn the user that the given matrix is not a valid one (for example, a 2x2 correlation matrix never has "large" non diagonal elements). Currently, the code does not complain when it gets an invalid matrix.

[ces42]

#103 affects some of this. The code in the pull request raises an error if the covariance/correlation matrices are not positive semidefinite (e. g. off-diagonal elements larger than one in the covariance matrix). Symmetry is not enforced since only the lower half of the matrices is read (I think this is probably the best option for this case?).
Also a correlation matrix is not forced to have ones on the diagonal, and if this isn't the case correlated_values_norm won't even return values with the prescribed uncertainty. IMO that last part should be changed.