Radiological Sign (RS) datasets need refactoring

The way support for _rs splits was implemented required a special arrangement in ptbench. I'm proposing we just re-structure this as yet another "experiment" in which the input database-splits contain the full dataset (i.e., the radiological sign predictions from a multi-class classifier). We then simply create a simple pytorch LogisticRegression model we use with these. Everything else should follow the same design pattern as we have for direct image-to-probability setups.