WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild