cuda::atomic failures with underaligned types

[jrhemstad]

The following code will fail if the default alignment (4) is used for key.

#if 1
struct key {
  int32_t a;
  int32_t b;
};
#else
struct alignas(8) key {
  int32_t a;
  int32_t b;
};
#endif
int main(){
    static_assert(alignof(key) == 4);
    cuda::atomic<key> k;
    auto r = k.load();
}

Error:

/opt/compiler-explorer/libs/libcudacxx/trunk/include/cuda/std/detail/__atomic(224): error: no suitable user-defined conversion from "cuda::std::__3::__cxx_atomic_alignment_wrapper_t<key>" to "key" exists
          detected during:
            instantiation of "_Tp cuda::std::__3::__cxx_atomic_load(const volatile cuda::std::__3::__cxx_atomic_base_impl_default<_Tp, _Sco> *, int) [with _Tp=key, _Sco=0]" 
/opt/compiler-explorer/libs/libcudacxx/trunk/include/cuda/std/detail/libcxx/include/atomic(1665): here
            instantiation of "_Tp cuda::std::__3::__atomic_base<_Tp, _Sco, <unnamed>>::load(cuda::std::__3::memory_order) const noexcept [with _Tp=key, _Sco=0, <unnamed>=false]" 

Specifying a larger alignment (8) resolves the issue.

https://godbolt.org/z/ETohMxqWb

[wmaxey]

I added a test to address this in #179

https://github.com/NVIDIA/libcudacxx/pull/179/files#diff-847d589c3abbc6908eec536e2d7551d25ef63f93f81df0a10b9d62b5595af325

[wmaxey]

Closing with the caveat that this still fails for Pascal devices, but rather than a compile time error it fails at runtime with unsupported shared/global memory operation.