Transfer learning-assisted inverse modeling in nanophotonics based on mixture density networks