LLM-based Translation Inference with Iterative Bilingual Understanding