Transfer Learning Across Heterogeneous Features For Efficient Tensor Program Generation