Self-Supervised Learning for Multimodal Non-Rigid 3D Shape Matching