Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions