Synthesizing Long-Term Human Motions with Diffusion Models via Coherent Sampling