ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models