Towards Geospatial Foundation Models via Continual Pretraining