Artifacts and Attention Sinks: Structured Approximations for Vision Transformers