Improve testing for DecisionTree variances

The current test assumes that Impurity.calculate() returns the variance correctly. A better approach would be to test if the variance returned equals the variance that we can manually verify on a toy data and tree.