Multimodal Pretraining, Adaptation, and Generation for Recommendation: A Survey