Quantum advantage for learning shallow neural networks with natural data distributions