MAMO: Masked Multimodal Modeling for Fine-Grained Vision-Language Representation Learning