Block Sparse Matrices for Smaller and Faster Language Models