[libcxx] Threading support: Externalize hardware_concurrency()

[rmaprath]

Summary

Another one of those platform-dependent methods which should live behind the threading API.

[rmaprath]

Re-based on trunk.

[rmaprath]

Ping?

[EricWF]

Same comment as D29818

I really don't like this change (and D29818) because they start to include a ton of extra headers, and because they lift complex configuration logic into what should be a very simple header.
These functions are much better suited to being define out-of-line for a number of reasons.

Is there another way we can allow "external implementations" of these functions without lifting everything into a header?

[rmaprath]

Different take on the patch: Externalize this function only for the externally-thread-api variant.

This is much less intrusive. Note that I haven't added the declaration of __libcpp_thread_hw_concurrency() into __threading_support because it doesn't belong there (needs to be provided through a custom __external_threading header instead).

There is no easy way to test this apart from building an actual external-thread-api libc++ variant. We could do some form of testing with the external-thread-library configuration, but it still requires some not-so-pretty changes that is probably best avoided.

[rmaprath]

@EricWF: Ping?

src/thread.cpp

thread::detach()
 
     if (ec)
         __throw_system_error(ec, "thread::detach failed");
 }
 
 unsigned
 thread::hardware_concurrency() _NOEXCEPT
 {
-#if defined(CTL_HW) && defined(HW_NCPU)
+#if defined(_LIBCPP_HAS_THREAD_API_EXTERNAL)
+    // Defer to the external threading implementation
+    return __libcpp_thread_hw_concurrency();
+#elif defined(CTL_HW) && defined(HW_NCPU)
     unsigned n;
     int mib[2] = {CTL_HW, HW_NCPU};
     std::size_t s = sizeof(n);
     sysctl(mib, 2, &n, &s, 0, 0);
     return n;
 #elif defined(_SC_NPROCESSORS_ONLN)
     long result = sysconf(_SC_NPROCESSORS_ONLN);
     // sysconf returns -1 if the name is invalid, the option does not exist or