Hardness Results for Neural Network Approximation Problems