Kokkos Atomics in v3.6.01

[Yurlungur]

I recently discovered that on our local Cray cluster, I need to build with

-DKokkos_ENABLE_IMPL_DESUL_ATOMICS=OFF

or I encounter several difficulties:

• In downstream codes with relocatable device code, Kokkos::initialize fails immediately
• Even with no downstream RDC, with cuda-aware MPI, a code path eventually triggers, which causes an error.

In both cases, the error looks like

terminate called after throwing an instance of 'std::runtime_error'
  what():  Desul::Error: init_lock_arrays_cuda: post init kernel error(cudaErrorIllegalAddress): an illegal memory access was encountered

I've reported this issue to the Kokkos team. I'm not sure if we need to do anything about this, but I think it's worth documenting for posterity.