auto regression does not return the *co*variance

[mathause]

Ok, this is a bit akward, but what I did in #309 was wrong - not in that it leads to a wrong result, but the name is wrong. The AR process returns the variance and not the covariance!

What I found confusing is that once this variance is used to draw auto regression time series, and once the co-variance from another source is used. I should try to make that clearer.

[mathause]

Ok all this confusion is self made :-( Originally there were two code paths for drawing AR samples

• Before refactoring:
  • uncorrelated samples -> used np.random.normal -> required std
  • correlated samples -> used np.random.multivariate_normal -> required cov
• After refactoring:
  • uncorrelated samples -> uses np.random.multivariate_normal -> requires 1D cov
  • correlated samples -> uses np.random.multivariate_normal -> requires 2D cov

I am very tempted to just revert and have _fit_auto_regression_xr return std. Especially because I am thinking about splitting the code path again (for performance reasons, see #165).

However, as we average the variability of the residuals (see #307) its better to return them as var. I think...