Support atomics with memory orders

[bernhardmgruber]

AFAIU the atomics that alpaka exhibt a relaxed memory order. That is, performing an atomic operation in alpaka does neither synchronize with other atomic operatons, nor give any guarantee on the visibility of preceding/subsequent ordinary loads and stores. We effectively assume a weakly-ordered memory model. Any synchchronization of loads and stores requires the use of explicit memory fences in addition to relaxed atomic operations. This is not how C++ does atomics since C++11.

In C++11 (and C11, Rust, etc.) atomic operations can be configured on their effect on the visible order of memory operations. This configuraiton is the memory order. CUDA's libc++ started to offer such configurable atomics a while back, see for example the definition of atomic_ref. I have not found a similar feature in HIP yet. SYCL has those: https://registry.khronos.org/SYCL/specs/sycl-2020/html/sycl-2020.html#_memory_ordering

For the efficient implementation of certain synchronization primitives, e.g. locks, or certain lock-free data structures, e.g. lock-free queues, atomics with various memory orders are a necessity. To that end, I wonder when and how those should be supported in alpaka. I am mostly leaning towards a class type similar to std::atomic_ref, but I am open to suggestions.