remove unsupported unsafe-atomic flag

The latest ROCm compiler doesn't seem to support it

CMakeLists.txt

@@ -63,8 +63,6 @@ set(GINKGO_CUDA_COMPILER_FLAGS "" CACHE STRING
     "Set the required NVCC compiler flags, mainly used for warnings. Current default is an empty string")
 set(GINKGO_CUDA_ARCHITECTURES "Auto" CACHE STRING
     "A list of target NVIDIA GPU architectures. See README.md for more detail.")
-# the details of fine/coarse grain memory and unsafe atomic are available https://docs.olcf.ornl.gov/systems/crusher_quick_start_guide.html#floating-point-fp-atomic-operations-and-coarse-fine-grained-memory-allocations
-option(GINKGO_HIP_AMD_UNSAFE_ATOMIC "Compiler uses unsafe floating point atomic (only for AMD GPU and ROCM >= 5). Default is ON because we use hipMalloc, which is always on coarse grain. Must turn off when allocating memory on fine grain" ON)
 set(GINKGO_HIP_COMPILER_FLAGS "" CACHE STRING
     "Set the required HIP compiler flags. Current default is an empty string.")
 set(GINKGO_HIP_NVCC_COMPILER_FLAGS "" CACHE STRING

cmake/hip.cmake

@@ -206,9 +206,6 @@ endif()
 set(GINKGO_HIPCC_OPTIONS ${GINKGO_HIP_COMPILER_FLAGS} "-std=c++14 -DGKO_COMPILING_HIP")
 set(GINKGO_HIP_NVCC_OPTIONS ${GINKGO_HIP_NVCC_COMPILER_FLAGS} ${GINKGO_HIP_NVCC_ARCH} ${GINKGO_HIP_NVCC_ADDITIONAL_FLAGS})
 set(GINKGO_HIP_CLANG_OPTIONS ${GINKGO_HIP_CLANG_COMPILER_FLAGS} ${GINKGO_AMD_ARCH_FLAGS})
-if(GINKGO_HIP_AMD_UNSAFE_ATOMIC AND HIP_VERSION VERSION_GREATER_EQUAL 5)
-    list(APPEND GINKGO_HIP_CLANG_OPTIONS -munsafe-fp-atomics)
-endif()
 # HIP's cmake support secretly carries around global state to remember
 # whether we created any shared libraries, and sets PIC flags accordingly.
 # CMake's scoping rules means that this makes the hip_add_* calls order- and