Logger: find the NVTX3 headers next to nvcc too, so the aarch64 cross build compiles

CUDA::nvtx3 is only an alias for CUDAToolkit_INCLUDE_DIRS. In the aarch64 SBSA cross build that is
targets/sbsa-linux/include, and the cross-sbsa packages the image installs (cudart, crt, culibos,
cufft) do not put nvtx3 there, so Logger.cpp failed with "nvtx3/nvToolsExt.h: No such file or
directory". NVTX v3 is header-only and architecture-neutral, so the headers are now looked for in the
toolkit include dirs and then in the include dir beside nvcc (the host tree, which ships them).
One CMake decision sets JFJOCH_USE_NVTX on JFJochLogger; where no headers are found the ranges
compile out, as they already do without CUDA.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

common/CMakeLists.txt

@@ -43,8 +43,21 @@ ADD_LIBRARY(JFJochLogger STATIC
 )
 
 TARGET_LINK_LIBRARIES(JFJochLogger PUBLIC spdlog::spdlog JFJochVersion)
+# NVTX v3 is header-only and architecture-neutral. CUDA::nvtx3 is no help here: it only points at
+# CUDAToolkit_INCLUDE_DIRS, which in a cross build is the target tree (targets/sbsa-linux/include),
+# and NVIDIA's cross-sbsa packages do not put nvtx3 there. So look in the target tree first and then
+# in the toolkit's own include dir next to nvcc; where neither has it, the ranges compile out.
 IF (JFJOCH_CUDA_AVAILABLE)
-    TARGET_LINK_LIBRARIES(JFJochLogger PRIVATE CUDA::nvtx3)   # header-only NVTX v3
+    FIND_PATH(JFJOCH_NVTX3_INCLUDE_DIR nvtx3/nvToolsExt.h
+            HINTS ${CUDAToolkit_INCLUDE_DIRS} "${CUDAToolkit_BIN_DIR}/../include"
+            NO_DEFAULT_PATH NO_CMAKE_FIND_ROOT_PATH)
+    IF (JFJOCH_NVTX3_INCLUDE_DIR)
+        TARGET_INCLUDE_DIRECTORIES(JFJochLogger SYSTEM PRIVATE "${JFJOCH_NVTX3_INCLUDE_DIR}")
+        TARGET_LINK_LIBRARIES(JFJochLogger PRIVATE ${CMAKE_DL_LIBS})   # NVTX loads its injection library with dlopen
+        TARGET_COMPILE_DEFINITIONS(JFJochLogger PRIVATE JFJOCH_USE_NVTX)
+    ELSE()
+        MESSAGE(STATUS "NVTX3 headers not found - NVTX ranges compiled out")
+    ENDIF()
 ENDIF()
 
 ADD_LIBRARY(JFJochZMQ STATIC ZMQWrappers.cpp ZMQWrappers.h)

common/Logger.cpp

@@ -9,7 +9,7 @@
 #include "Logger.h"
 #include "GitInfo.h"
 
-#ifdef JFJOCH_USE_CUDA
+#ifdef JFJOCH_USE_NVTX
 #include <nvtx3/nvToolsExt.h>
 
 uint64_t NvtxRangeStart(const std::string &name) {