Approximation bounds for convolutional neural networks in operator learning