Developing Japanese CLIP Models Leveraging an Open-weight LLM for Large-scale Dataset Translation