1. Would it be worthwhile to add parameters to control missingness in dataset generators?

   I need this for benchmarking [MRG] ENH Add support for missing values to Tree based Classifiers #5974. Thought this might come in handy for teaching too.

   Typically I would like to add missing_rate, target_correlation_rate, feature_correlation_rate and missing_values.

   The target_correlation_rate would control the extent to which the dataset is MNAR^* and feature_correlation_rate would control the extent to which the dataset is MAR^†.

   target_correlation_rate + feature_correlation_rate <= 1

   1 - (target_correlation_rate + feature_correlation_rate) would control the extent to which the dataset is MCAR^‡.

   Does this sound good?

2. Either as an addition or as an alternative to 1, could we have missing transformers with the above described params?

* - Missing Not At Random (Missingness is correlated with the target)

† - Missing At Random but correlated with the other feature values.

‡ - Missing Completely At Random (No correlation with either the target or parameters)

Ping @agramfort @glouppe @GaelVaroquaux