Adaptive Search Control Through Meta-Gradient Reinforcement Learning