Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model