Optimal Approximation and Learning Rates for Deep Convolutional Neural Networks