UNIMO-3: Multi-granularity Interaction for Vision-Language Representation Learning