Training Long-Context, Multi-Turn Software Engineering Agents with Reinforcement Learning