Consider polar projections in Raster spatial join bounding box calculation

Filing this as an issue so we don't forget!

> Extent reprojection requires internal points across the whole domain, not just the boundary. I can demonstrate if desired, my approach is in reproj::reproj_extent

_Originally posted by @mdsumner in [#1126](https://github.com/apache/sedona-db/pull/1126/changes/13f834a1ee048e97f894ba016bb25e50a0e1c87b..497e36bdfa8b7f757fb3b49cf653006e81adb1f4#r3732660054)_

This is a great point...in addition to the example in the linked thread, any polar projection when converted to lat/lon falls in this category (e.g., the south pole is directly in the middle of some projections, so even densifying the edges is not going to work, and even including points in the middle is not going to necessarily catch the minimum). Some of the minimum values it covers might also not be coverable by the other projection (e.g., mercator projections with a sigularity at the south pole).

Probably a geography covering is more accurate than a bounding box.

Start with the reproj::reproj_extent approach in the linked discussion and locate the current Raster spatial join bounding-box calculation. Work through polar projections, interior extrema, and cases where a bounding box is insufficient, then define regression coverage for the demonstrated examples. No source or test file is named in the issue.

Written by the indexing model from the issue text.