Improve specificity function

[MichaelRoeder]

The search for invariant expressions uses a function g that is defined as the harmonic mean of the variance-inspired measure h' for all positive graph examples and 1 minus the negative graph examples. h' summarizes the quality of lambda across the negative examples. This might be not so good since it may hide that lambda is an invariant for a lot of the negative examples - but not for all. Maybe extending 'g' to have several runs of h' for each of the negative example types (i.e., for the set of graphs that come from the same graph generator). This should prefer expressions that are invariant in non of the negative examples.