Characterizing higher-order representations through generative diffusion models explains human decoded neurofeedback performance