Repro - Do We Need Adam? Surprisingly Strong and Sparse Reinforcement Learning with SGD in LLMs