[RF][CMake] Allow building RooBatchCompute_CUDA standalone

[guitargeek]

The CUDA backend of RooBatchCompute is the only part of ROOT that needs the CUDA toolkit, and it is a plugin: libRooBatchCompute never links against it, it loads it by name at runtime from RooBatchCompute::initCUDA() with gSystem->Load("libRooBatchCompute_CUDA"). There is therefore no technical reason why it has to be built together with the rest of ROOT, and packagers (conda-forge in particular) would like to ship a CUDA-free ROOT plus a separate package that only provides the GPU kernels. So far that required carrying a hand-written CMakeLists.txt downstream.

roofit/batchcompute/CMakeLists.txt now doubles as the top-level CMakeLists.txt of a standalone project that builds nothing but libRooBatchCompute_CUDA:

  cmake -S <root-src>/roofit/batchcompute -B build \
        -DCMAKE_PREFIX_PATH=<root-prefix>
  cmake --build build
  cmake --install build

When the file is processed as the top-level list file it sets up its own project(), finds the installed ROOT, enables the CUDA language and skips the part that defines libRooBatchCompute and the CPU backends. The definition of the CUDA target itself is shared between the two modes, so the list of source files exists only once. In standalone mode the target is created with a plain add_library() rather than ROOT_LINKER_LIBRARY(), because the latter relies on variables that only exist inside a ROOT build tree; it links against the imported ROOT::RooBatchCompute target, which owns the dispatchCUDA pointer that the library overwrites when it is loaded, and it takes the private headers straight from the source tree since res/ is not part of the ROOT installation. Like ROOT_LINKER_LIBRARY() it forces the "lib" prefix, which on Windows is not what CMake would pick by default but is the name gSystem->Load() asks for, and like the ROOT build it defaults CMAKE_BUILD_TYPE to Release. A ROOT installation that has no RooFit is rejected with an explicit error, and one that already ships libRooBatchCompute_CUDA gets a warning that the default install destination would overwrite it.

By default the library is installed into the library directory of the ROOT installation it was configured against, which is where the dispatcher looks for it; RooBatchCompute_CUDA_INSTALL_LIBDIR overrides that and accepts a relative path so that the usual CMAKE_INSTALL_PREFIX/DESTDIR staging works.

The regular ROOT build is unaffected: -Dcuda=ON still builds the CUDA backend exactly as before.

Closes #18641

🤖 Done with the help of AI

[hageboeck]

Good idea! I might have found a few improvements.

[github-actions]

Test Results

    22 files      22 suites   3d 13h 17m 33s ⏱️
 3 862 tests  3 861 ✅ 0 💤 1 ❌
75 843 runs  75 842 ✅ 0 💤 1 ❌

For more details on these failures, see this check.

Results for commit 66cdea1.

[guitargeek]

Thank you @hageboeck for your review! Your comments all made sense to me, so I have implemented the suggestions one-to-one.