CoFB: latency-constrained co-scheduling of flows and batches for deep learning inference service on the CPU–GPU system