SINGER: Vivid Audio-driven Singing Video Generation with Multi-scale Spectral Diffusion Model