Hybrid Contrastive Learning of Tri-Modal Representation for Multimodal Sentiment Analysis