1 Boomerang Distillation Enables Zero-Shot Model Size Interpolation Harvard Data Centric Machine Learning Group 19