Variance-based acceptance only works correctly for rules that match all inputs

[heidmic]

https://github.com/heidmic/suprb/blob/2af04212216ffd6a01160167f50ee01e62b2b36c/suprb/optimizer/rule/acceptance.py#L27

@dpaetzel says: if mean(local_y) == 0, then this is a 1/beta–weighted variance, however, this assumption is wrong for non-general rules