SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution