A Scalable MIP-based Method for Learning Optimal Multivariate Decision Trees