Learning Functional Graphs with Nonlinear Sufficient Dimension Reduction