Disentangling 3D Prototypical Networks for Few-Shot Concept Learning