Statistics for Efficient Inference of Transformers in Natural Language Processing: Early Exiting and Beyond