Statistical Guarantees for Approximate Stationary Points of Shallow Neural Networks