PRIME: Scaffolding Manipulation Tasks with Behavior Primitives for Data-Efficient Imitation Learning