Migrate to `cuda::std::atomic` when possible

[brycelelbach]

[alliepiper]

Making note of some internal conversations:

This is currently not feasible because libcu++ atomics aren't available on several architectures that Thrust/CUB support. The only use of an atomic operation in NVBench is in blocking_kernel, and it is encapsulated behind an API that doesn't expose any details about the atomic API used.

This is low priority since it would drop pre-Volta support with no real benefit, and will likely be backlogged until the CTK-supported-arches and the libcudacxx-atomic-supported-arches are better aligned.