Locality Guidance for Improving Vision Transformers on Tiny Datasets