WIP: update CUDA support

[white238]

This updates BLT's CUDA support from the deprecated FindCUDA path to CMake's
built-in CUDA language and CUDAToolkit support.

Changes

• Replace FindCUDA with native CUDA language enablement and find_package(CUDAToolkit REQUIRED).
• Map the BLT CUDA runtime target to CMake's CUDA::cudart target.
• Define blt::cuda_runtime as depending on CUDA::cudart, and blt::cuda as depending on cuda_runtime, matching the HIP target mapping style.
• Remove Clang CUDA support, including ENABLE_CLANG_CUDA, BLT_CLANG_CUDA_ARCH, related setup/config propagation, and the dedicated cuda-arch test project.
• Remove legacy CUDA link-rule overrides that were needed for FindCUDA, including custom CMAKE_CUDA_LINK_EXECUTABLE/device-link rules and CUDA rpath flag overrides.
• Use CUDAToolkit_ROOT as the CUDA toolkit root variable. CUDA_TOOLKIT_ROOT_DIR is only read as a compatibility input and mapped to CUDAToolkit_ROOT.
• Move separable compilation support to native CMake behavior via CMAKE_CUDA_SEPARABLE_COMPILATION. The old CUDA_SEPARABLE_COMPILATION variable is only mapped for compatibility.
• Update CUDA host-config examples and tutorial text for CUDAToolkit_ROOT and native CMake CUDA support.
• Fix CUDA/MPI test compatibility with hwloc 2.x. The build exposed an error where HWLOC_TOPOLOGY_FLAG_WHOLE_IO was undefined; the test now uses hwloc_topology_set_io_types_filter(..., HWLOC_TYPE_FILTER_KEEP_ALL) for hwloc 2.x and keeps HWLOC_TOPOLOGY_FLAG_WHOLE_IO for older hwloc versions.
• Fix a Clang CUDA warning in the CUDA/MPI test caused by a variable-length array:
  cudaStream_t cuStreams[cudaStreams]
  The test now uses a single cudaStream_t for the local rank and destroys it during cleanup.
• Update CUDA documentation after reviewing docs/tutorial/common_hpc_dependencies.rst and the rest of docs/. The tutorial now describes blt::cuda as compiling sources as CUDA language sources, and blt::cuda_runtime as adding CUDA runtime usage without changing source language. Removed stale CUDA_LINK_WITH_NVCC documentation and updated object-library CUDA device-link wording for native CMake CUDA behavior.
• Add internal CUDA compile coverage for the migrated target behavior: t_cuda_runtime_compile verifies blt::cuda_runtime provides CUDA runtime headers/libraries without compiling the source as CUDA, and t_cuda_language_compile verifies blt::cuda changes a .cpp source to CUDA language compilation.