Hybrid Stochastic-Binary Computing for Low-Latency and High-Precision Inference of CNNs