Fast Distributed Deep Learning via Worker-adaptive Batch Sizing