ArtiFree: Detecting and Reducing Generative Artifacts in Diffusion-based Speech Enhancement