Forecast-MAE: Self-supervised Pre-training for Motion Forecasting with Masked Autoencoders