Tunable Subnetwork Splitting for Model-parallelism of Neural Network Training