Quantized Neural Network Inference with Precision Batching