Add overview (pyramid) generation for Cloud-Optimized GeoTIFF output

[brendancol]

to_geotiff() can write GeoTIFFs, but production Cloud-Optimized GeoTIFFs need internal overviews (image pyramids) for efficient tiled access at multiple zoom levels. Without overviews, COGs written by xarray-spatial still need a post-processing step through GDAL (gdaladdo) to be truly usable, which defeats the purpose of having native I/O.

Scope

• Generate 2x downsampled overviews and embed them in the output GeoTIFF.
• Configurable number of overview levels, or auto-calculate based on raster dimensions (keep halving until the smallest overview fits in a single tile).
• Resampling method per overview level:
  • nearest for categorical data
  • average for continuous data
  • mode for majority-class downsampling
  • Similar to what gdaladdo provides.
• Integrate with the existing to_geotiff() / write_geotiff_gpu() pipeline so users get proper COGs in one call instead of writing a flat TIFF and running a separate tool.

Current workaround

Write a GeoTIFF with xarray-spatial, then shell out to gdaladdo and gdal_translate -of COG to add overviews and restructure. This adds a hard GDAL dependency and creates intermediate files.

[brendancol]

This is implemented on master. The to_geotiff() COG writer supports overview_levels (auto-calculated when omitted) and overview_resampling (mean, nearest, min, max, median, mode, cubic). Overviews are generated inline during COG writes with no GDAL post-processing step needed. See xrspatial/geotiff/_writer.py and tests in test_cog.py / test_features.py.