Improving MLLM's Document Image Machine Translation via Synchronously Self-reviewing Its OCR Proficiency