SCANET: Improving multimodal representation and fusion with sparse‐ and cross‐attention for multimodal sentiment analysis | Synapse