Lottery Tickets on a Data Diet: Finding Initializations with Sparse Trainable Networks