SPIRE: Synergistic Planning, Imitation, and Reinforcement Learning for Long-Horizon Manipulation