Revealing the Power of Spatial-Temporal Masked Autoencoders in Multivariate Time Series Forecasting