Systematically include data observations marked as manually edited

In the context of multishot learning, training dataset balancing rules that involve subsampling majority classes may dismiss manual corrections made on top of automatic predictions.

Following up on feature #93 (closed), data observations with secondary label "edited" should be systematically included as representatives of the associated primary label(s). Random subsampling would still be use to complete the classes up to the desired sample sizes.

This will be the new default behavior in combination with the auto and maggotuba balancing rules.