SEEKR: Selective Attention-Guided Knowledge Retention for Continual Learning of Large Language Models