<generator>: Various fixes and cleanups

[StephanTLavavej]

My apologies for the grab-bag PR (slightly excusable for a feature branch).

This starts with trivial followups to #4464, then addresses pre-existing issues that I found:

• std:: => _STD
• Include <cstddef> for std::nullptr_t.
• Extract a verbose immediately-invoked lambda for clarity.
• Bugfix: Use __stdcall to make <generator> immune to default calling conventions.
  • We must always do this when mentioning function pointer types in headers.
• Scalar constexpr size_t _Align doesn't need to be static.
• Fix unintentional move from const _Alloc&.
  • This is now consistent with the other occurrence of auto& _Stored_al.
• Consistently use _RANGES.
• _Promise => _CoroPromise to avoid intolerable shadowing of <future>.
  • I clearly missed this in the shipping <coroutine>. We conventionally never allow template parameters to shadow other machinery like this:

    STL/stl/inc/future

    Lines 1001 to 1002 in d6efe94

    	template <class _Ty>
    	class _Promise {

• throw; => _RERAISE;
  • We never directly try/catch/throw in headers. Using the macros respects _HAS_EXCEPTIONS=0.
• Overhaul generator's static_asserts.
  • Add Standard citations.
  • "generator's second argument" was bogus; use "generator's selected reference type" to describe italic-monospace-reference. Then say "an actual reference type" for clarity.
  • "an iterator with [...] cannot model indirectly_readable" was confusing because it was describing a bad situation, but the other messages are describing how situations must be good. Rephrase this for clarity.
• Mark _Dealloc_fn as _NOEXCEPT_FNPTR.
  • This is important because the compiler can't see through function pointers.
• Use static member functions instead of lambdas to make Clang happy.
  • Clang -fno-ms-compatibility refuses to convert stateless lambdas to __stdcall function pointers, which I suppose is a defensible choice.