Robust Multi-View Learning via Representation Fusion of Sample-Level Attention and Alignment of Simulated Perturbation