Reinvestigate minimum required CMake version for different architectures.

[kennyweiss]

We recently updated Axom's C++ language requirement to c++20, which has some potential implications on the CMake requirements on various supported architectures. Let's reinvestigate the minimum required CMake versions for our supported CPU and GPU architectures.

Among other concerns, this potential impacts:

• our stated version in the cmake_minimum_required call, which impacts our available CMake policies and their defaults
  

  axom/src/CMakeLists.txt

  Lines 10 to 20 in 97f1e49

  	if (ENABLE_HIP OR AXOM_ENABLE_HIP)
  	cmake_minimum_required(VERSION 3.21)
  	else()
  	# Do not bump this over 3.14 due to it changing CMake policies and breaking
  	# the CUDA build. We check the required version specifically for this below
  	cmake_minimum_required(VERSION 3.14)
  	endif()
  	# Default unset third-party CMP0074 handling to NEW so imported packages honor
  	# <PackageName>_ROOT hints without emitting developer warnings.
  	set(CMAKE_POLICY_DEFAULT_CMP0074 NEW CACHE STRING "")

• our version checks for CUDA platforms
  

  axom/src/CMakeLists.txt

  Lines 134 to 137 in 97f1e49

  	# Check for minimum CMake version required w/o changing policies like cmake_minimum_required
  	if(AXOM_ENABLE_CUDA AND ${CMAKE_VERSION} VERSION_LESS 3.18.0)
  	message(FATAL_ERROR "Axom requires CMake version 3.18.0+ when CUDA is enabled.")
  	endif()

• our "getting started" documentation
  

  axom/src/docs/sphinx/quickstart_guide/config_build.rst

  Lines 26 to 33 in 97f1e49

  	Basic requirements:
  	~~~~~~~~~~~~~~~~~~~
  	* C++ compiler with C++17 support at a minimum
  	* CMake with a minimum required version of 3.14 for CPU-only and CUDA builds,
  	and a minimum version of 3.21 when building with HIP support
  	* Fortran Compiler (optional)
  	* Python 3.8 at a minimum for python bindings (optional)

  
  (note that we are updating this section for C++20 in Some touchups to reflect that C++20 is now required for Axom #1957)

This might also be affected by a BLT PR to improve CUDA support: llnl/blt#769

See also, the discussion on a recent docs update PR: #1957 (comment)
In particular, @white238 wrote:

CMake 3.18 + CUDA + c++20 actually is a failure due to CMake not knowing that CUDA supports C++20 that early in CMake. It was set to this because many moons ago we noticed that it was changing cmake policies based on the cmake_minimum_required and this caused failures. We should reevaluate this at some point.