DeepSeek-VL: Towards Real-World Vision-Language Understanding