One Step at a Time: Pros and Cons of Multi-Step Meta-Gradient Reinforcement Learning