Higher-Order Function Networks for Learning Composable 3D Object Representations