All You Need Is Supervised Learning: From Imitation Learning to Meta-RL With Upside Down RL