Improve compile times

[ericniebler]

• Apply "SCARY iterator"-like refactorization of class templates to minimize the number and complexity of class template arguments.
• No local classes in function templates. That includes lambdas.
• Bring down mangled name length to speed up debug builds.
• Investigate where complicated function template return types are being computed unnecessarily. See https://godbolt.org/z/Mqhj7voxq for an example. See cwg2369 for the core issue.
• Use Templight to get a compile-time flame graph and prune expensive template instantiations.
• Replace internal use of heavyweight types (std::tuple, std::optional, std::variant) with more lean custom types. See manual_lifetime and manual_lifetime_union in libunifex. See my prototype lambda-based tuple implementation in this gist.