SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training