improve clang-cuda CI jobs

* clang-10 doesn't work with HIP since nvcc doesn't like clang > 9
* we need to propagate the CUDA compiler to the installation,
otherwise tiny mismatches in the handling of constexpr globals
cause link errors.

.gitlab-ci.yml

@@ -383,7 +383,6 @@ build/clang-cuda101/gcc/all/release/shared:
     CUDA_COMPILER: "clang++"
     BUILD_OMP: "ON"
     BUILD_CUDA: "ON"
-    BUILD_HIP: "ON"
     BUILD_TYPE: "Release"
     CUDA_ARCH: 35
   tags:
@@ -401,7 +400,6 @@ build/clang-cuda101/clang/all/release/shared:
     CUDA_COMPILER: "clang++"
     BUILD_OMP: "ON"
     BUILD_CUDA: "ON"
-    BUILD_HIP: "ON"
     BUILD_TYPE: "Release"
     BUILD_SHARED_LIBS: "OFF"
     CUDA_ARCH: 35

cmake/GinkgoConfig.cmake.in

@@ -135,7 +135,13 @@ include(${CMAKE_CURRENT_LIST_DIR}/windows_helpers.cmake)
 # NOTE: we do not export benchmarks, examples, tests or devel tools
 #     so `third_party` libraries are currently unneeded.
 
-# populate CUDA_HOST_COMPILER if Ginkgo was built with CUDA
+# propagate CUDA_COMPILER if Ginkgo was built with CUDA
+if (GINKGO_BUILD_CUDA AND NOT CMAKE_CUDA_COMPILER)
+    message(STATUS "Ginkgo: Setting CUDA compiler to ${GINKGO_CUDA_COMPILER}")
+    set(CMAKE_CUDA_COMPILER "${GINKGO_CUDA_COMPILER}" CACHE STRING "" FORCE)
+endif()
+
+# propagate CUDA_HOST_COMPILER if Ginkgo was built with CUDA
 if (GINKGO_BUILD_CUDA AND GINKGO_CUDA_HOST_COMPILER AND NOT CMAKE_CUDA_HOST_COMPILER)
     message(STATUS "Ginkgo: Setting CUDA host compiler to ${GINKGO_CXX_COMPILER}")
     set(CMAKE_CUDA_HOST_COMPILER "${GINKGO_CXX_COMPILER}" CACHE STRING "" FORCE)