TinyLLaVA: A Framework of Small-scale Large Multimodal Models