Evaluating Large Language Models for Document-grounded Response Generation in Information-Seeking Dialogues