AutoCFR: Learning to Design Counterfactual Regret Minimization Algorithms