Thinking Before Looking: Improving Multimodal LLM Reasoning via Mitigating Visual Hallucination