Lotus: Diffusion-based Visual Foundation Model for High-quality Dense Prediction