Evaluating Large Vision-and-Language Models on Children's Mathematical Olympiads