Add feature level argument to limone for "expected sign"

[taylorlarsen]

In the previous implementation of limone there was a "up is good?" argument that ensured that impactability guidance supported the use case such that if "up is good?" was set to "NO", it would only surface changes in features that decreased the predicted probability.

The above functionality is helpful, but it would be great if the future implementation of limone (patient impact predictor) had an additional feature level argument for "expected sign" to ensure that impactability predictions are consistent with best practice and the literature and that no guidance is offered if the direction of the insight is not as expected.

An example of this might be bathing compliance for the inpatient central line population. The literature and best practice supports that a higher bathing rate is associated with lower risk of a central line associated blood stream infection (CLABSI). Regardless of what the model might predict, it would be important to have a safeguard in place that ensures that limone only offers guidance that an increase in a patient's bathing rate decreases the predicted probability of a CLABSI. In this example, it would be important to prevent surfacing guidance that suggests decreasing a patient's bathing rate will decrease the predicted probability of a CLABSI.

[michaellevy]

This will be for numerics only. If a categorical variable is provided, error with reference to excluded_transitions