An Inference Perspective on Meta-Reinforcement Learning