A Careful Examination of Large Language Model Performance on Grade School Arithmetic