Leveraging Diffusion Disentangled Representations to Mitigate Shortcuts in Underspecified Visual Tasks