FastVLM: Efficient vision encoding for vision language models