FusionStitching: Boosting Execution Efficiency of Memory Intensive Computations for DL Workloads