Counting Ability of Large Language Models and Impact of Tokenization