Add atomic intrinsics

[tkf]

This patch introduces LLVM.Interop.atomic_pointer* on LLVMPtr mimicking Core.Intrinsics.atomic_pointer* API on Ptr. This is a superset of JuliaGPU/CUDA.jl#1393. It extracts the "CUDA-unaware" part of the atomics previously partially implemented in CUDA.jl. My motivation for this is to implement GPU-interop of https://github.com/JuliaConcurrent/UnsafeAtomics.jl and https://github.com/JuliaConcurrent/Atomix.jl without explicitly dispatching on CUDA.jl and AMDGPU.jl types.

As discussed in JuliaGPU/CUDA.jl#1393 the native code from this is not particularly good on CUDA. But that's probably fixable in LLVM (a long-term solution). As a short-term solution, I think CUDA.jl and AMDGPU.jl can use @device_override to override these intrinsics in their device context (e.g., use @llvm.nvvm.* intrinsics in CUDA).

cc @vchuravy @jpsamaroo

[vchuravy]

This is complex enough that having it in one place seems sensible.

[vchuravy]

1.7 failed on the atomics test

[maleadt]

I'm not really familiar with atomics and their APIs, but superficially this looks OK (minor some small changes).

CI sees to fail with assertions enabled though.

[tkf]

@maleadt I'm fixing bugs locally with LLVM assertion enabled. I'll ping you once I fix all bugs locally.

[tkf]

@maleadt OK, I think it's ready. Can you run the CI?