Pretrained Diffusion Models for Unified Human Motion Synthesis