Major changes
* use Woodbury identity for grouped random effects models -> faster inference for grouped random effects / mixed effects models
* no profiling out of error variance for Fisher scoring -> faster learning of covariance paramters
* add functionality that negative log-likelihood can be evaluated