Low Rank and Sparse Fourier Structure in Recurrent Networks Trained on Modular Addition