Tight Sample Complexity of Learning One-hidden-layer Convolutional Neural Networks