Improve the modularity of the *ColumnStatsMerger classes

*ColumnStatsMerger classes contain a lot of duplicate code which is not specific to the data type, and that could therefore be lifted to a common parent class.

This phenomenon is bound to become even worse if we keep enriching further our supported set of statistics as we did in the context of HIVE-26221.

The current ticket aims at improving the modularity and code reuse of the *ColumnStatsMerger classes, while improving unit-test coverage to cover all classes and support more use-cases.