SAIL: Self-supervised Learning of Lighting-Invariant Representations from Real Images with Latent Diffusion