Minimax Regret Optimization for Robust Machine Learning under Distribution Shift