radeonsi: always create wait_mem_scratch for compute-based culling