Fix clang CI build: use clang as nvcc host compiler (#6004)

## Summary
Fix CI `clang-build-23` job failure where nvcc used gcc as host compiler
but received clang-specific flags (`-fclang-abi-compat=17`) from
PyTorch's CMake config.

Two issues were fixed:
- **`setup-env.sh`**: Switch from `CC="ccache clang"` to `CC=clang` +
`CMAKE_*_COMPILER_LAUNCHER=ccache`. The old pattern caused CMake to
resolve the CUDA host compiler (`-ccbin`) to `ccache` instead of clang.
Set `CUDAHOSTCXX` to tell CMake to use clang++ as nvcc's host compiler.
- **`CMakeLists.txt`**: Read `CUDAHOSTCXX` into
`CMAKE_CUDA_HOST_COMPILER` before `enable_language(CUDA)` to ensure it
takes effect before PyTorch's TorchConfig.cmake can override it.

## Test plan
- [x] CI `clang-build-23` job passes

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: xwang233

CMakeLists.txt

@@ -4,6 +4,14 @@
 cmake_minimum_required(VERSION 3.18)
 project(nvfuser)
 
+# When CXX is clang, ensure nvcc also uses clang as its host compiler.
+# This must be set before enable_language(CUDA) so that CMake configures
+# nvcc with the correct -ccbin. The CUDAHOSTCXX env var should do this,
+# but PyTorch's TorchConfig.cmake can override it, so we also set it here.
+if(NOT CMAKE_CUDA_HOST_COMPILER AND DEFINED ENV{CUDAHOSTCXX})
+  set(CMAKE_CUDA_HOST_COMPILER $ENV{CUDAHOSTCXX})
+endif()
+
 enable_language(CUDA)
 
 cmake_policy(SET CMP0063 NEW) # make symbol visibility always apply

tools/setup-env.sh

@@ -21,15 +21,23 @@ else
     exit 1
 fi
 
-# Use ccache if available and CC/CXX are not already set
+# Set CC/CXX to the raw compiler paths. Use CMAKE_<LANG>_COMPILER_LAUNCHER
+# for ccache instead of CC="ccache clang", because the "ccache clang" form
+# causes CMake (and PyTorch's TorchConfig.cmake) to resolve the CUDA host
+# compiler (-ccbin) to ccache instead of clang, breaking nvcc.
 if [ -z "$CC" ]; then
+    export CC=$CLANG_PATH
+    export CXX=$CLANGXX_PATH
     if command -v ccache >/dev/null 2>&1; then
-        export CC="ccache $CLANG_PATH"
-        export CXX="ccache $CLANGXX_PATH"
-    else
-        export CC=$CLANG_PATH
-        export CXX=$CLANGXX_PATH
+        export CMAKE_C_COMPILER_LAUNCHER=ccache
+        export CMAKE_CXX_COMPILER_LAUNCHER=ccache
+        export CMAKE_CUDA_COMPILER_LAUNCHER=ccache
     fi
 fi
 
+# Tell nvcc to use clang++ as its host compiler. Without this, nvcc defaults
+# to gcc, which fails on clang-specific flags like -fclang-abi-compat that
+# PyTorch adds to CMAKE_CXX_FLAGS.
+export CUDAHOSTCXX=$CLANGXX_PATH
+
 export TORCH_CUDA_ARCH_LIST="10.0"