Optimizing Stochastic Computing for Low Latency Inference of Convolutional Neural Networks