fix handling of standard_error_threshold in feature selection

[cleboo]

Hi!

Love the package - really useful!

I noticed a bug in the get_reduced_features_set method of ShapRFECV when best_coherent or best_parsimonious is used: it does not actually take the variance of the evaluation metrics into account. Instead it uses the value of standard_error_threshold itself to set a lower bound for the metric when selecting the number of features to use.

This pull request fixes that.

Best wishes and thanks for sharing your package!

[cleboo]

I added another fix of the logic for "best" feature selection. The logic was actually the wrong way around (checking if the best value lies within the variation around the current value, instead of checking if the current value is inside the range around the best value).