Diffusion Model with Cross Attention as an Inductive Bias for Disentanglement