All Languages Matter: On the Multilingual Safety of Large Language Models