Deconstructing Denoising Diffusion Models for Self-Supervised Learning