Prevent usage of cuda::atomic::fetch_max/min for non-integral types.

[wmaxey]

I've used a static_assert instead of adding a proper inheritance layer. This should serve as a temporary stop-gap and provide meaningful output for users if they attempt to use fetch_max/min on non-integral types.

This doesn't seem to impact other units and the output is somewhat understandable.

C:/sbf/libcudacxx/include\cuda/std/atomic(111): error: static assertion failed with "fetch_max/min is available for integral types only"
          detected during:
            instantiation of "_Tp cuda::__4::atomic<_Tp, _Sco>::fetch_min(const _Tp &, cuda::__4::memory_order) volatile noexcept [with _Tp=float, _Sco=cuda::std::__4::__detail::thread_scope_system]"
C:\sbf\libcudacxx\.upstream-tests\test\cuda\atomics\atomic.ext\atomic_fetch.fail.cpp(30): here
            instantiation of "void TestFn<T, Selector, <unnamed>>::operator()() const [with T=float, Selector=local_memory_selector, <unnamed>=cuda::std::__4::__detail::thread_scope_thread]"
C:\sbf\libcudacxx\.upstream-tests\test\cuda\atomics\atomic.ext\atomic_fetch.fail.cpp(53): here