Hidden Progress in Deep Learning: SGD Learns Parities Near the Computational Limit