Toolset update: VS 2022 17.3 Preview 2, CUDA 11.6.0

[StephanTLavavej]

This updates the toolset to:

• VS 2022 17.3 Preview 2.
  • Required in the README, but not in <yvals_core.h> because the internal build still uses 19.32.
• CMake 3.23.
  • Required in the README and CMakeLists.txt.
• Python 3.10.5.
  • We still require Python 3.9 because the VS Installer still has 3.9.7.
• CUDA 11.6.0.
  • azure-devops/provision-image.ps1 now installs all CUDA features, which is much easier than enumerating almost all of them.
  • Required in <yvals_core.h>.
    • CUDA 11.6.0 was the first release that officially supported VS 2022, so this aligns our support.
    • Approved by our boss and double boss, who are awesome! 😸

Then we can remove workarounds:

• tests/std/tests/GH_000639_nvcc_include_all/env.lst
  • Drop --allow-unsupported-compiler, now that we're using a CUDA-supported scenario.
• tests/std/tests/P1502R1_standard_library_header_units/test.cpp
  • Remove workarounds for VSO-1496084 "Standard Library Header Units: Splitting <chrono> emits fatal error C1116: unrecoverable error importing header unit '<chrono>'".
• Remove workarounds for VSO-1269037 "Compiler ICE on reference to pointer in constexpr".
  • <vector> 2051-2056: This is clearer with the workaround as _Temp is used repeatedly, but we can extract it.
  • <vector> 3282: This is clearer with the workaround as _Temp is used repeatedly.
  • <xmemory> 1212: Remove the workaround as _Temp is used only once.
  • <xmemory> 1233-1236: The original code used pointers-to-pointers, which were difficult to read. The workaround was significantly clearer, but still had subtle behavior - it was writing to auto& _Pnext repeatedly, ultimately nulling it out (so the separate assignment was redundant). As @CaseyCarter observed, we can avoid the subtlety by using exchange() to null out _Myproxy->_Myfirstiter immediately, and then the rest of the loop can use a non-reference variable.
• <format>
  • Remove workaround for VSO-1444086 "consteval + /presetPadding error".
• Remove workarounds for internal test machines running older OSes.
  • There is a slight chance that I'll need to drop this (I am informed that the PR/CI system upgrade to "1ES Phase 5" makes this unnecessary, but we may still want/need to retain this for the internal Contest pool).
• <yvals_core.h>
  • Remove the _MSVC_CONSTEXPR fallback definition. This is related to Annotate some functions _MSVC_CONSTEXPR #2658, where MSVC-PR-391437 added this to VCRuntime.
• CMakeLists.txt
  • Change /d1FastFail to /fastfail. This is documented by cl /?, and was added in VS 2019 16.1 by MSVC-PR-162175.

I am planning a followup PR to address #1301 by removing _HAS_CONDITIONAL_EXPLICIT now that CUDA 11.6.0 supports it; this should be separate because the change is large and may need to be reverted.

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.