GeoUtils Distance accuracy degrades with irregular rectangles [LUCENE-6994]

This is a continuation of #7966 which validates USGS accuracy of 0.5% for Sinnott haversine distance calculation. The issue also introduced a space segmentation approach for BKD distance queries near the poles. In #8013 a restriction on distance size was initially removed to randomly test BKD distance queries at any range. This revealed an issue where distance error nearly doubles for exotic rectangles created by BKD's split algorithm. This issue will investigate potential distance error caused by the segmentation approach introduced in the second part of #7966.

---
Migrated from [LUCENE-6994](https://issues.apache.org/jira/browse/LUCENE-6994) by Nick Knize (@nknize), updated Feb 01 2016

Start by reading #7966 and #8013, then inspect the GeoUtils distance calculation and BKD distance-query space segmentation described here. Reproduce the randomized queries that exposed nearly doubled error on exotic rectangles and determine whether segmentation causes it; done means the distance error is explained and accuracy is validated against the stated 0.5% USGS target.

Written by the indexing model from the issue text.