Use thread_local instead of __thread under OS X

stout's THREAD_LOCAL was introduced as an abstraction since Apple's C++ toolchain didn't support thread_local at the time. It uses the extension __thread under OS X, so we cannot create non-POD thread-local vars.

With Xcode-8 (i.e., with Sierra) and above Apple's toolchain does support thread_local, and we should eventually e.g., remove the THREAD_LOCAL abstraction.