LFM2-VL: Efficient Vision-Language Models