Coordinating Distributed Example Orders for Provably Accelerated Training