Mini-Gemini: Mining the Potential of Multi-modality Vision Language Models