Hardness of Noise-Free Learning for Two-Hidden-Layer Neural Networks