We need to change #include <nvToolsExt.h> into #include <nvtx3/nvToolsExt.h>.

The old form (#include <nvToolsExt.h>) requires linking to dynamic library (libnvToolsExt.so), while the new form is header-only. In PR #548, @cliffburdick changed CMakeLists.txt to make matx link to CUDA::nvtx3 instead of UDA::nvToolsExt if CMake has a higher version that supports it. Thus if you use a CMake version >= 3.25, you will only use nvtx3 headers, without linking to libnvToolsExt.so. But we are still including the old <nvToolsExt.h>, rather than the newer <nvtx3/nvToolsExt.h>. For some functions, the old <nvToolsExt.h> don't contain implementations. This makes the linking fail with undefined reference errors.

For example, using CMake 3.28, I cannot build the conv2d target in examples. The linker outputs:

[ 50%] Building CUDA object examples/CMakeFiles/conv2d.dir/conv2d.cu.o
[100%] Linking CUDA executable conv2d
/usr/bin/ld: CMakeFiles/conv2d.dir/conv2d.cu.o: in function `matx::endEvent(int)':
/workspaces/MatX/include/matx/core/nvtx.h:189: undefined reference to `nvtxRangeEnd'
collect2: error: ld returned 1 exit status
make[3]: *** [examples/CMakeFiles/conv2d.dir/build.make:110: examples/conv2d] Error 1
make[2]: *** [CMakeFiles/Makefile2:316: examples/CMakeFiles/conv2d.dir/all] Error 2
make[1]: *** [CMakeFiles/Makefile2:323: examples/CMakeFiles/conv2d.dir/rule] Error 2
make: *** [Makefile:255: conv2d] Error 2

To fix this error, we can change to use #include <nvtx3/nvToolsExt.h>. It's also recommended by NVTX's README:

NOTE: Older versions of NVTX did require linking against a dynamic library. NVTX version 3 provides the same API, but removes the need to link with any library. Ensure you are including NVTX v3 by using the nvtx3 directory as a prefix in your #includes:
#include <nvtx3/nvToolsExt.h>
...