Embedding spatial context in urban traffic forecasting with contrastive pre-training