Create a tinyllava with qwen2-0.5B-instruct

[bil-ash]

Please create a tinyllava with Qwen2-0.5B-instruct and Florence-2-base (I guess that the finetuned version is not suitable for VLM). It will become the smallest VLM because till now the smallest one is nanoLLAVA. Along with model checkpoints, also please release the training code (python files for the custom LLM and visual encoder) .