High-dimensional covariance matrix estimation using a low-rank and diagonal decomposition