large distance errors appearing in the short-range case by Andoyer and Thomas formulas

Hello,
The Boost Geometry's distance calculations by Andoyer and Thomas formulas seem to induce large errors in the short-range case, as shown in
Figure 1. The maximum absolute distance error as a function of distance. in [GEODESIC ALGORITHMS: AN EXPERIMENTAL STUDY](https://pdfs.semanticscholar.org/0fd3/7bed6be199ee1766ae46a6ec2ed409d0304c.pdf) by Vissarion Fisikopoulos (2019).

I think they should use [law of haversines](https://en.wikipedia.org/wiki/Haversine_formula#The_law_of_haversines) formula for distance (i.e., d = 2 * asin(sin_d_half)) for accuracy, instead of [spherical law of cosines](https://en.wikipedia.org/wiki/Spherical_law_of_cosines) (i.e., d = acos(cos_d)).
Thank you.

Start by locating the Boost.Geometry implementations of the Andoyer and Thomas distance formulas and their existing distance tests. Reproduce the short-range errors described in the issue and compare the current spherical-law-of-cosines calculation with the proposed haversine approach. Done means the short-range accuracy is improved and regression coverage is added.

Written by the indexing model from the issue text.