A Flexible Sparsity-Aware Accelerator with High Sensitivity and Efficient Operation for Convolutional Neural Networks