Probing LLM World Models: Enhancing Guesstimation with Wisdom of Crowds Decoding