The Hubbard model beyond the two-pole approximation: a composite operator method study