Deal with negative estimated standard deviation
In #392 we had the issue that we do not enforce positive standard deviations. I thus proposed 72208b1b. @stepan-a suggested that optimizers might dislike the resulting cliff at 0 and that negative standard deviations are perfectly equivalent for the likelihood to the positive ones. Unfortunately, I don't see this as they are not always squared for covariance elements. In any case, if @stepan-a is correct, we still need to adapt the saving and displaying of parameter estimation results to not confuse users with negative numbers.