Generate output (preds) in common geospatial formats

[allanjust]

Potentially use gdal to generate multilayer cloud-optimized geotiff by year (UTC). One layer per overpass over the prediction region.

• overpass time was tile dependent (with ~5 min variation across adjacent tiles) -- perhaps will need to pick a reference tile?
• unclear what other metadata to pass through
• sometimes helpful to have original (raw) data for contrast with predicted values (pred); is this just useful for our evaluation papers or should we export both values for others or as separate (optional) files?

[allanjust]

This looks helpful:
https://geoexamples.com/other/2019-02-08-cog-tutorial/

[allanjust]

perhaps we could consider parquet as an efficient target store for the prediction output (as big tabular data) still within the targets workflow (and more language agnostic than fst);
targets appears to leverage arrow::read_parquet() and arrow::write_parquet()