[libc++] Casey's P/R for LWG3664 "LWG3392 broke std::ranges::distance(a, a+3)"

[ldionne]

Summary

This seems equivalent to what we've got today, but doesn't re-evaluate sized_sentinel_for, so it's arguably easier on the compiler.
Casting int(&)[10] to int* const& strikes me as "clever", but I think it works.

Don't merge until LWG3664 is resolved, or at least seems likely to resolve in this direction. But let's poke buildkite with this and prove that it doesn't fail any of our existing regression tests.

The corresponding PR for Microsoft STL is https://github.com/microsoft/STL/pull/2522

[ldionne]

[Github PR transition cleanup]

Dropping since this is incomplete (at least missing tests).

libcxx/include/__iterator/distance.h

while (__first != __last) {
       ++__first;
       ++__n;
     }
     return __n;
   }
 
   template<class _Ip, sized_sentinel_for<decay_t<_Ip>> _Sp>
   _LIBCPP_HIDE_FROM_ABI
-  constexpr iter_difference_t<_Ip> operator()(_Ip&& __first, _Sp __last) const {
-    if constexpr (sized_sentinel_for<_Sp, __uncvref_t<_Ip>>) {
-      return __last - __first;
-    } else {
-      return __last - decay_t<_Ip>(__first);
-    }
+  constexpr iter_difference_t<decay_t<_Ip>> operator()(_Ip&& __first, _Sp __last) const {
+    return __last - static_cast<const decay_t<_Ip>&>(__first);
   }
 
   template<range _Rp>
   _LIBCPP_HIDE_FROM_ABI
   constexpr range_difference_t<_Rp> operator()(_Rp&& __r) const {
     if constexpr (sized_range<_Rp>) {
       return static_cast<range_difference_t<_Rp>>(ranges::size(__r));
     } else {