Optimal block designs for three treatments when observations are correlated