Simplify numerical scales

[JohelEGP]

• Change term of non-linear scale to numerical scale. Suggested in https://github.com/nholthaus/units/issues/185#issuecomment-423107386.
• Add linearized_value_t tag to disambiguate the unit constructor whose value is already linearized.
• Reduce the scales' scope of functionality to what their name suggest. Make them non-template entities with only a linearize and scale static member functions in place of their old constructor and operator() respectively. Move m_value and operator() to unit, where they naturally belong.
• unit still inherits from NumericalScale, but privately. This is so that the scale traits continue to work.
• Rename m_value to linearized_value, upgrading its explaining comment to code. [1] [2]
• Drop constexpr support from decibel_scale, which uses non-constexpr std math functions.
• Remove has_value_member, now unused.

With this, the fact that the unit values are stored linearized has effectively become an implementation detail, I think.

I think this is desirable as the current scales can be abused because they're publicly bases of unit, their constructors were overly accepting, and the non-linear scale concept is underspecified.

[nholthaus]

great change. I had in my head at one point that the scales could support really whacky things (like richter scale or other offset scales similar to how temperature is handled), but it was very poorly named and pointlessly complex in the end, in addition to never being exercised.