Robust Multimodal Learning for Ophthalmic Disease Grading via Disentangled Representation