LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches