On the Fundamental Limits of Coded Data Shuffling for Distributed Machine Learning