LVLM-COUNT: Enhancing the Counting Ability of Large Vision-Language Models