Draft-and-Revise: Effective Image Generation with Contextual RQ-Transformer