Turing completeness is a red herring and not an argument against using a language. Turing completeness is usually practically useful only to prove that certain analyses are impossible.

As to why, sometimes you want the flexibility of inferring settings from third-party data sources, or implicitly from the layout of the directory structure, or naming conventions, or any other thing you choose to mention. IMO the power of a programmable configuration language is that the configuration need not actually be written in the language, but stored somewhere else, and just retrieved using the language.