Spotify-confidence

------------------

* Fixed bug that led to crash when a segment had zero observations

------------------

* Added support for variance reduction using linear regression

------------------

* Added SampleSizeCalculator class to allow for more complex sample size calculations involving several metrics and dimensions

------------------

* Improve performance by using more optimal pandas operations and by running computations in paralell over groupby dimensions.

------------------

* Bugfix: When you had some metrics with NIMs and some with MDEs, one of them overrid the other
* Switching to using N-1 instead of N in denominator of variance estimate to get unbiased estimator for smaller sample sizes. For binary metrics we still use the old formula, equivalent to p*(1-p).
* Removing powered_effect_metric column because it's identical to powered_effect after bugfix in 2.4.1
* Minor performace and robustness tweaks to sequential bounds solver.

------------------

* Bugfix: The field "powered_effect_for_metric" in the output of the difference methods (when verbose=True) was computed using incorrect current_number_of_units