BRAINTEASER: Lateral Thinking Puzzles for Large Language Models