Partition-Based Formulations for Mixed-Integer Optimization of Trained ReLU Neural Networks