Add scoped atomic_thread_fence

[vchuravy]

As noticed in EnzymeAD/Enzyme.jl#511 CUDA C++ support a wider selection of memory orders
and emits different assembly for SM_70 and above: https://godbolt.org/z/Y7Pj5G7sK

For now I just added the memory fences necessary to implement the rest.

@tkf @maleadt over the long-term I would be in favor of moving to Atomix.jl instead of CUDA.@atomic
is there any shared infrastructure we can use? As you see I am defining scope and order here again.

[codecov]

Codecov Report

Base: 61.68% // Head: 60.08% // Decreases project coverage by -1.60% ⚠️

Coverage data is based on head (654870d) compared to base (dcb175e).
Patch has no changes to coverable lines.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1644      +/-   ##
==========================================
- Coverage   61.68%   60.08%   -1.61%     
==========================================
  Files         151      151              
  Lines       11349    10833     -516     
==========================================
- Hits         7001     6509     -492     
+ Misses       4348     4324      -24     

Impacted Files	Coverage Δ
lib/cusolver/linalg.jl	49.71% <0.00%> (-36.73%)	⬇️
lib/cublas/CUBLAS.jl	52.25% <0.00%> (-24.07%)	⬇️
lib/cusparse/conversions.jl	79.77% <0.00%> (-14.31%)	⬇️
lib/cusparse/interfaces.jl	58.55% <0.00%> (-13.77%)	⬇️
src/compiler/gpucompiler.jl	82.14% <0.00%> (-11.20%)	⬇️
lib/cusparse/level3.jl	63.85% <0.00%> (-10.85%)	⬇️
lib/cusparse/broadcast.jl	25.92% <0.00%> (-10.35%)	⬇️
lib/cusparse/types.jl	41.89% <0.00%> (-8.11%)	⬇️
lib/cusparse/generic.jl	89.30% <0.00%> (-7.87%)	⬇️
src/utilities.jl	68.57% <0.00%> (-7.75%)	⬇️
... and 71 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[jgreener64]

Does this need something else before it can be merged as an intermediate solution? I am finding it useful to get Atomix.@atomic :monotonic working with Enzyme on GPU (JuliaConcurrent/Atomix.jl#33 and EnzymeAD/Enzyme.jl#511).

[vchuravy]

This shouldn't work with Enzyme, since Enzyme won't understand the assembly inserted, that's one of the reasons I haven't pushed on this further.

[jgreener64]

It seems to work with EnzymeAD/Enzyme.jl#511 and in another context I tried, unless I am getting something mixed up or you mean it will only work in specific cases.