DAPT: A Dual Attention Framework for Parameter-Efficient Continual Learning of Large Language Models