Set flag to always use unsafe atomic storage (#18590)

Closes #18587

This PR sets `_LIBCUDACXX_ATOMIC_UNSAFE_AUTOMATIC_STORAGE` as a workaround for NVIDIA/cccl#4586. This will unblock Spark-RAPIDS CI.

Authors:
  - Yunsong Wang (https://github.com/PointKernel)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - Jihoon Son (https://github.com/jihoonson)
  - Nghia Truong (https://github.com/ttnghia)

URL: #18590

Author: PointKernel

cpp/CMakeLists.txt

@@ -939,7 +939,11 @@ target_compile_definitions(cudf PRIVATE $<$<BOOL:${CUDF_KVIKIO_REMOTE_IO}>:CUDF_
 
 # Remove this after upgrading to a CCCL that has a proper CMake option. See
 # https://github.com/NVIDIA/cccl/pull/2844
-target_compile_definitions(cudf PRIVATE THRUST_FORCE_32_BIT_OFFSET_TYPE=1 CCCL_AVOID_SORT_UNROLL=1)
+target_compile_definitions(
+  cudf
+  PUBLIC _LIBCUDACXX_ATOMIC_UNSAFE_AUTOMATIC_STORAGE=1
+  PRIVATE THRUST_FORCE_32_BIT_OFFSET_TYPE=1 CCCL_AVOID_SORT_UNROLL=1
+)
 
 # Compile stringified JIT sources first
 add_dependencies(cudf jitify_preprocess_run)

cpp/benchmarks/CMakeLists.txt

@@ -98,7 +98,9 @@ function(ConfigureBench CMAKE_BENCH_NAME)
                CUDA_STANDARD_REQUIRED ON
   )
   target_compile_definitions(
-    ${CMAKE_BENCH_NAME} PRIVATE THRUST_FORCE_32_BIT_OFFSET_TYPE=1 CCCL_AVOID_SORT_UNROLL=1
+    ${CMAKE_BENCH_NAME}
+    PUBLIC _LIBCUDACXX_ATOMIC_UNSAFE_AUTOMATIC_STORAGE=1
+    PRIVATE THRUST_FORCE_32_BIT_OFFSET_TYPE=1 CCCL_AVOID_SORT_UNROLL=1
   )
   target_link_libraries(
     ${CMAKE_BENCH_NAME} PRIVATE cudf_benchmark_common cudf_datagen benchmark::benchmark_main

cpp/tests/CMakeLists.txt

@@ -59,7 +59,9 @@ function(ConfigureTest CMAKE_TEST_NAME)
   )
 
   target_compile_definitions(
-    ${CMAKE_TEST_NAME} PRIVATE THRUST_FORCE_32_BIT_OFFSET_TYPE=1 CCCL_AVOID_SORT_UNROLL=1
+    ${CMAKE_TEST_NAME}
+    PUBLIC _LIBCUDACXX_ATOMIC_UNSAFE_AUTOMATIC_STORAGE=1
+    PRIVATE THRUST_FORCE_32_BIT_OFFSET_TYPE=1 CCCL_AVOID_SORT_UNROLL=1
   )
 
   target_link_libraries(