FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective