We-Math: Does Your Large Multimodal Model Achieve Human-like Mathematical Reasoning?