Propagate CuTe DSL cudart shim and --wrap to static-lib consumers

cute_dsl_setup() linked the cutedsl cudart shim and the CUDA<12.8
-Wl,--wrap=_cudaLaunchKernelEx option PRIVATEly. For STATIC_LIBRARY link
targets a static archive performs no link step, so neither reached the final
executable, breaking the CuTe DSL kernels on JetPack 6 / CUDA 12.6.

Propagate the shim via PUBLIC and the --wrap link option via INTERFACE for
static-library targets so the consuming executable inherits them; keep PRIVATE
for shared/other targets.

Signed-off-by: suharvest <suharvest@gmail.com>

Author: suharvest

cmake/CuteDsl.cmake

@@ -775,7 +775,8 @@ function(cute_dsl_setup)
   foreach(_tgt ${ARG_LINK_TARGETS})
     get_target_property(_tgt_type ${_tgt} TYPE)
     if(_tgt_type STREQUAL "STATIC_LIBRARY")
-      target_link_libraries(${_tgt} PUBLIC "${_static_lib}")
+      target_link_libraries(${_tgt} PUBLIC "${_static_lib}"
+                                           trt_edgellm_cutedsl_cudart_shim)
     else()
       target_link_libraries(${_tgt} PRIVATE "${_static_lib}"
                                             trt_edgellm_cutedsl_cudart_shim)
@@ -787,7 +788,11 @@ function(cute_dsl_setup)
     # JetPack 6).
     if(NOT _cute_dsl_cuda_ver STREQUAL "" AND _cute_dsl_cuda_ver VERSION_LESS
                                               12.8)
-      target_link_options(${_tgt} PRIVATE "-Wl,--wrap=_cudaLaunchKernelEx")
+      if(_tgt_type STREQUAL "STATIC_LIBRARY")
+        target_link_options(${_tgt} INTERFACE "-Wl,--wrap=_cudaLaunchKernelEx")
+      else()
+        target_link_options(${_tgt} PRIVATE "-Wl,--wrap=_cudaLaunchKernelEx")
+      endif()
     endif()
   endforeach()