Correct issues from investigation of VSO-938163

[CaseyCarter]

I've been helping the MSVC compiler team's constexpr specialist investigate VSO-938163 (which we'll call "constexpr hates ranges::iter_swap") and its interactions with our Ranges tests. She discovered there are really two different bugs that we've been calling VSO-938163, and fixed them both. In the process we found some miscategorization and other problems hidden behind VSO-938163.

• <algorithm>: Applied perma-workaround for VSO-1141368 ("constexpr hates lifetime extension") by avoiding structured binding and the consequent lifetime extension of temporaries implicitly bound to references.

• tests/std/tests/P0896R4_P1614R2_comparisons: This test triggers a different bug for which I filed VSO-1168721 with a minimal repro:

  struct B {};
  struct D : B {};
  
  constexpr bool test() {
      D some_deriveds[2] = {};
  
      return &some_deriveds[0] < static_cast<B *>(&some_deriveds[1]);
  }
  static_assert(test(), "FAIL");

  I've changed the TRANSITION comment appropriately.

• tests/std/tests/P0896R4_ranges_alg_partition: In addition to VSO-938163, this triggers VSO-1168723 in which std::array<std::pair<int, int>, N>'s implicit copy assignment operator is incorrectly not considered constexpr despite that std::pair<int, int> can be copy assigned in a constant expression. I've applied a perma-workaround to avoid VSO-1168723.

• tests/std/tests/P0896R4_ranges_iterator_machinery: Change #ifdef __clang__ guard to #if defined(__clang__) || defined(__EDG__) for consistency.

I've verified that the tests will all pass with the fixed compiler after applying these changes and removing all remaining workarounds for VSO-938163, which we'll be able to do after the fixed compiler ships.

[StephanTLavavej]

Thanks for investigating and cleaning up these workarounds! 😺 This helps keep the STL from accumulating unnecessary complexity.