Migrate to libcu++ atomics

This is blocked on NVIDIA/libcudacxx#52 and NVIDIA/cccl#946.

NVIDIA/libcudacxx#52 means we can just update the callsites, and not juggle multiple types (with different APIs and semantics, depending on SM version and operating system) through multiple layers of dispatch / agent level abstractions. This will lead to a far more maintainable, much lower effort implementation.