Evaluating the Elementary Multilingual Capabilities of Large Language Models with MultiQ