Measuring Multimodal Mathematical Reasoning with MATH-Vision Dataset