Coordinated Batching and DVFS for DNN Inference on GPU Accelerators