An Efficient Hardware Accelerator for Sparse Transformer Neural Networks