FastVLM: Efficient Vision Encoding for Vision Language Models