use CMake builtin mechanisms for setting target CUDA architecture(s)

[jameslamb]

Contributes to #115

As described at https://cmake.org/cmake/help/latest/variable/CMAKE_CUDA_ARCHITECTURES.html, if the target GPU architectures isn't explicitly provided to CMake, it falls back to a concrete (but system-and-compiler-specific) version.

To help with that, #81 introduced two things:

• ability to set the target architecture via an environment variable CUDA_ARCH
• fallback to native (whatever GPU is detected on the system where the build is running)

This proposes preserving that functionality, but in a different way:

• using the environment variable CUDAARCHS, which CMake has supported since v3.18 (docs link), instead of introducing a custom one
• ensuring that architecture can also be overridden for pip install

Notes for Reviewers

Benefits of these changes

Reduces complexity in managing CUDA architecture, by reducing the number of paths through which configuration flows to CMake. This will be important when we introduce yet another build pattern (conda builds) here.

How I tested this

Added code like this at the bottom of the top-level CMakeLists.txt.

get_target_property(
  REQUESTED_ARCHITECTURES
  legateboost
  CUDA_ARCHITECTURES
)

message(FATAL_ERROR "requested architectures: ${REQUESTED_ARCHITECTURES}")

Found the expected behavior for regular CMake builds, build.sh, and pip install.

[jameslamb]

I especially want your opinion on this @trivialfis, to be sure I correctly understood the goals of #81

[trivialfis]

Looks good to me! The PR I created was to prefer the CMake standard argument over the original custom argument. Anything that removes magic is welcome!

[jameslamb]

Great, thank you!