Augmenting Reinforcement Learning with Behavior Primitives for Diverse Manipulation Tasks