XTRUST: On the Multilingual Trustworthiness of Large Language Models