fitted.values in non-mixed null models are not what we expect

[amstilp]

The fitted.values element of null models is different for non-mixed models and mixed-models. This had caused errors in SPA p-values for models without variance components or where variance components were estimated to be 0 (see #84).

Consider a bigger fix to make fitted.values consistent for all types of null model (e.g., it is always the same thing) but that is a much more involved change and has some issues with backwards compatibility (e.g., using a version of the null model run with an older version of GENESIS to do Score.SPA association tests in a new version).

Originally posted by @amstilp in #84 (comment)