Improve GeoPointDistanceQuery performance [LUCENE-7663]

GeoPoint queries currently use only the bounding box for filtering.
But the query circle is only roughly 80% of the bounding box, so we could be roughly 20% faster. Furthermore, the current approach requires splitting the box if it crosses the date line.

> Schubert, E., Zimek, A., & Kriegel, H. P. (2013, August). Geodetic distance queries on r-trees for indexing geographic data. In International Symposium on Spatial and Temporal Databases (pp. 146-164).

The minimum spherical distance of a point to a rectangle is given ("Algorithm 2: Optimized Minimum Distance Point to MBR"). Rectangles whose minimum distance is larger than the query radius can be skipped. The authors used the R-tree, but it will work with any bounding box, so it can be used in CellComparator#relate.
It's not very complex - a few case distinctions, and then either Haversine distance, or cross-track-distance. So the cost ist comparable to Haversine.
This could be added as GeoRelationUtils.pointToRectMinimumDistance, for example.

This approach can also be used to priorize rectangles, for top-k search.

---
Migrated from [LUCENE-7663](https://issues.apache.org/jira/browse/LUCENE-7663) by Erich Schubert, updated Jan 28 2017

Start with GeoPointDistanceQuery and inspect CellComparator#relate, where bounding-box filtering is performed. Read the cited minimum-distance algorithm and consider GeoRelationUtils.pointToRectMinimumDistance; done means distance-aware rectangle filtering avoids unnecessary date-line splitting while preserving query results and performance.

Written by the indexing model from the issue text.