CMake: Default GPU Architectrues

[ptheywood]

When enabling CUDA (or HIP) as a language, CMake will either use a user provided value (via CMAKE_CUDA_ARCHITECTURES or env{CUDAARCHS}) or set CMAKE_CUDA_ARCHITECTURES to a defualt for the verison of CUDA used (50 for 12.x, 75 for 13.x).

At this point, we cannot tell if this value was set by the user or by CMake.

We currently provide a library-default of all-major, if this was not set by the user, via flamegpu_init_cuda_architectures which must be before CUDA is enabled, and then flamegpu_set_cuda_architectures is called either by the library target, or injected into the project command.

Although this is currently functional, adding in HIP support in #1379 significantly complicates this process, and if we were to add #260 and support find_package(flamegpu) would have some chicken-and-egg type issues related to languages and architecture flags.

If we remove our library-provided default and instead accept the CMake provided default, but strongly encourage users to build the correct architecture for their system we can:

• Not require users call flamegpu_init_cuda_architectures in their code (but maintain it as a deprecated noop for atleast a release)
• Delete a lot of complex CMake, which is hard to achieve consistent behaviour with for CUDA and AMD
• Just use all-major explicitily when required (and/or toolchain files)
• Avoid issues related to this when we add find_package(flamegpu) support.
• Potentially move some of flamegpu_set_cuda_architectures into a warning-generating function in case invalid values are set (although these errros may be caught by CUDA language support not being found in the first place)?

Approved by @mondus during FLAMEGPU team meeting

Todo:

• Replace the bodies of flamegpu_init_cuda_architectures and flamegpu_set_cuda_architectures with AUTHOR_WARNINGs stating that these methods are deprecated, no longer do anything and calls to them should be removed
• Remove internal calls to flamegpu_init_cuda_architectures in examples
• Potentially move some of flamegpu_set_cuda_architectures into a warning-generating function in case invalid values are set, although these errors will probably be caught by a try-compile failure
• Update CI to ensure CMAKE_CUDA_ARCHITECTURES is always used (this should be the case already)
• Update the Readme to strongly encourage the setting of CMAKE_CUDA_ARCHITECTURES/CUDAARCHS/CMAKE_HIP_ARCHITECTURES

[ptheywood]

The current explicit validation of CMAKE_CUDA_ARCHITECTURES is no longer applied anyway, atleast not with recent CMake which sets CMAKE_CUDA_ARCHITECTURES when enabling the CUDA language anyway, so the error message in that case is the same (and provides enough information to spot the issue)

CMake Error at /usr/share/cmake-3.28/Modules/CMakeTestCUDACompiler.cmake:59 (message):
  The CUDA compiler

    "/usr/local/cuda-13.0/bin/nvcc"

  is not able to compile a simple test program.

  It fails with the following output:

    Change Dir: '/home/ptheywood/code/flamegpu/FLAMEGPU2/build-cuarch-test/CMakeFiles/CMakeScratch/TryCompile-scwPgn'
    
    Run Build Command(s): /usr/bin/cmake -E env VERBOSE=1 /usr/bin/gmake -f Makefile cmTC_78716/fast
    /usr/bin/gmake  -f CMakeFiles/cmTC_78716.dir/build.make CMakeFiles/cmTC_78716.dir/build
    gmake[1]: Entering directory '/home/ptheywood/code/flamegpu/FLAMEGPU2/build-cuarch-test/CMakeFiles/CMakeScratch/TryCompile-scwPgn'
    Building CUDA object CMakeFiles/cmTC_78716.dir/main.cu.o
    /usr/local/cuda-13.0/bin/nvcc -forward-unknown-to-host-compiler -ccbin=/opt/lmod-modules/symlinks/gcc/14/g++   "--generate-code=arch=compute_12,code=[compute_12,sm_12]" -MD -MT CMakeFiles/cmTC_78716.dir/main.cu.o -MF CMakeFiles/cmTC_78716.dir/main.cu.o.d -x cu -c /home/ptheywood/code/flamegpu/FLAMEGPU2/build-cuarch-test/CMakeFiles/CMakeScratch/TryCompile-scwPgn/main.cu -o CMakeFiles/cmTC_78716.dir/main.cu.o
    nvcc fatal   : Unsupported gpu architecture 'compute_12'
    gmake[1]: *** [CMakeFiles/cmTC_78716.dir/build.make:79: CMakeFiles/cmTC_78716.dir/main.cu.o] Error 1
    gmake[1]: Leaving directory '/home/ptheywood/code/flamegpu/FLAMEGPU2/build-cuarch-test/CMakeFiles/CMakeScratch/TryCompile-scwPgn'
    gmake: *** [Makefile:127: cmTC_78716/fast] Error 2