Isometric Representation Learning for Disentangled Latent Space of Diffusion Models