MMEarth: Exploring Multi-Modal Pretext Tasks For Geospatial Representation Learning