Fix UB of potentially wrong compile-time assumption of transparently replaceable

[mingxwa]

Added necessary calls to std::launder() to avoid causing undefined behavior. No functional changes.

There is a discussion of std::launder() on stackoverflow. libc++ is also using std::launder() in its implementation of small buffer optimization.

Closes #80

[tian-lt-personal]

Suggest adding test cases to prevent from regressions.

[mingxwa]

Suggest adding test cases to prevent from regressions.

It does not seem to be easy to come up with a case that can break without std::launder() on GCC, clang or MSVC.

I checked the transmogrify() example on (cppreference)[https://en.cppreference.com/w/cpp/utility/launder]. It breaks on GCC and clang, but it is not how proxy works. I believe the behavior is generally undefined if a "transmogrify-able" type is used in an STL container or proxy.

Another way to attack is const fields. I tried many ways to trick the compilers, but none of the 3 compromises.

I also checked the unit tests of libstdc++. They have a case for std::launder(), but unfortunately the case passes even without std::launder() on recent GCC releases (I checked several versions with O2 and O3). Same as Microsoft implementation of STL.

Search for std::launder() in GitHub, I found a comment in abseil:

// A typed accessor for the object in `TypeErasedState` storage
template <class T>
T& ObjectInLocalStorage(TypeErasedState* const state) {
  // We launder here because the storage may be reused with the same type.
#if defined(__cpp_lib_launder) && __cpp_lib_launder >= 201606L
  return *std::launder(reinterpret_cast<T*>(&state->storage));
#elif ABSL_HAVE_BUILTIN(__builtin_launder)
  return *__builtin_launder(reinterpret_cast<T*>(&state->storage));
#else

  // When `std::launder` or equivalent are not available, we rely on undefined
  // behavior, which works as intended on Abseil's officially supported
  // platforms as of Q2 2022.
#if !defined(__clang__) && defined(__GNUC__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wstrict-aliasing"
#endif
  return *reinterpret_cast<T*>(&state->storage);
#if !defined(__clang__) && defined(__GNUC__)
#pragma GCC diagnostic pop
#endif

#endif
}

After research for a whole day, I do not think we should spend more time try to come up with a quality unit test case for this change.