Interaction-Aware Sampling-Based MPC with Learned Local Goal Predictions