A Divide-and-Conquer Method for Sparse Inverse Covariance Estimation