Learning Multiple Tasks with a Sparse Matrix-Normal Penalty