* Refactored the p-calculation out for easier parallelization

* Add a settings variable to steer the number of processes used for calculating the p-value significance

* Parallelize the pvalue calculation.

For this, the _calculate_p_values function needs to return a pd.Series and the
df_feature DataFrame must be updated with the content of these Series.

* Increased the coverage again by explicitely calling the _calculate_p_value function

* Added an explicit test for const features

* Always use all cpus

* Tried to fix the bug of too many cores in multiprocessing on travis.ci builds.