Demystifying Reinforcement Learning Post-Training of Language Models