A Scalable Attention-Based Approach for Image-to-3D Texture Mapping