ASTRAL: Automated Safety Testing of Large Language Models