TEaR: Improving LLM-based Machine Translation with Systematic Self-Refinement