Region-Aware Text-to-Image Generation via Hard Binding and Soft Refinement