From Dense to Dynamic: Token-Difficulty Driven MoEfication of Pre-Trained LLMs