Exploring Bit-Slice Sparsity in Deep Neural Networks for Efficient ReRAM-Based Deployment