From Regression to Classification: Exploring the Benefits of Categorical Representations of Energy in MLIPs