Naturalness of Attention: Revisiting Attention in Code Language Models