Add ordered atomics and enable KernelAbstractions atomic support

[maleadt]

Metal's low-level atomic intrinsics previously exposed only relaxed ordering. That is fine for counters and reductions, but not for synchronization patterns where ordinary device-memory writes are published through an atomic and then observed by another workitem.

Thread explicit memory orderings through the low-level atomic API and light up KernelAbstractions/Atomix atomics on the Metal backend:

• Add explicit order (and flags/scope) arguments to the low-level atomic store/load/exchange/fetch/compare-exchange intrinsics, using Val-specialized paths so orderings reach GPU code as compile-time constants.
• Keep the existing direct call sites on relaxed ordering, so code that only needs relaxed atomics keeps emitting the same air.atomic.* forms and does not change behaviour.
• Support independent success/failure orderings on compare-exchange.
• Gate the acquire/release/acq_rel/seq_cst orderings and the memory flags on Metal 4.1 with device-side @static_asserts (check_atomic_memory_order / check_atomic_flags), so unavailable operations fail to compile with a clear message and enclosing metal_version() guards can eliminate them.
• Enable KA.supports_atomics(::MetalBackend) when the device supports Metal 4.1, so Atomix-based kernels (e.g. KernelAbstractions' @atomic) work on Metal.

[codecov]

Codecov Report

❌ Patch coverage is 89.28571% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.26%. Comparing base (998b6cb) to head (7b392d3).
⚠️ Report is 4 commits behind head on main.

Files with missing lines	Patch %	Lines
src/version.jl	50.00%	5 Missing ⚠️
src/MetalKernels.jl	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #889      +/-   ##
==========================================
+ Coverage   85.93%   86.26%   +0.33%     
==========================================
  Files          76       76              
  Lines        5246     5300      +54     
==========================================
+ Hits         4508     4572      +64     
+ Misses        738      728      -10     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: 7b392d3	Previous: e527abe	Ratio
array/accumulate/Float32/1d	406042 ns	419167 ns	0.97
array/accumulate/Float32/dims=1	376458 ns	396959 ns	0.95
array/accumulate/Float32/dims=1L	8807208 ns	8882417 ns	0.99
array/accumulate/Float32/dims=2	446458 ns	455959 ns	0.98
array/accumulate/Float32/dims=2L	2594792 ns	2288250 ns	1.13
array/accumulate/Int64/1d	849167 ns	873792 ns	0.97
array/accumulate/Int64/dims=1	933750 ns	953833 ns	0.98
array/accumulate/Int64/dims=1L	9533416 ns	9673416 ns	0.99
array/accumulate/Int64/dims=2	1252500 ns	969791 ns	1.29
array/accumulate/Int64/dims=2L	6526667 ns	6737834 ns	0.97
array/broadcast	221125 ns	187875 ns	1.18
array/construct	2375 ns	2458 ns	0.97
array/permutedims/2d	453500 ns	431375 ns	1.05
array/permutedims/3d	1018083 ns	729667 ns	1.40
array/permutedims/4d	1142709 ns	1054333 ns	1.08
array/private/copy	236291 ns	227500 ns	1.04
array/private/copyto!/cpu_to_gpu	195667 ns	204500 ns	0.96
array/private/copyto!/gpu_to_cpu	190792 ns	203666 ns	0.94
array/private/copyto!/gpu_to_gpu	204208 ns	185917 ns	1.10
array/private/iteration/findall/bool	1059125 ns	1109917 ns	0.95
array/private/iteration/findall/int	1236625 ns	1243916 ns	0.99
array/private/iteration/findfirst/bool	1128625 ns	1088208 ns	1.04
array/private/iteration/findfirst/int	857917 ns	1118125 ns	0.77
array/private/iteration/findmin/1d	1230166 ns	1270125 ns	0.97
array/private/iteration/findmin/2d	1079666 ns	1056416 ns	1.02
array/private/iteration/logical	1547750 ns	1853875 ns	0.83
array/private/iteration/scalar	1231458 ns	1160458 ns	1.06
array/random/rand/Float32	435542 ns	406667 ns	1.07
array/random/rand/Int64	521459 ns	544667 ns	0.96
array/random/rand!/Float32	392291 ns	373875 ns	1.05
array/random/rand!/Int64	409292 ns	404125 ns	1.01
array/random/randn/Float32	392041 ns	395709 ns	0.99
array/random/randn!/Float32	351625 ns	331625 ns	1.06
array/reductions/mapreduce/Float32/1d	383791 ns	385083 ns	1.00
array/reductions/mapreduce/Float32/dims=1	339458 ns	325667 ns	1.04
array/reductions/mapreduce/Float32/dims=1L	626708 ns	631583 ns	0.99
array/reductions/mapreduce/Float32/dims=2	344750 ns	329250 ns	1.05
array/reductions/mapreduce/Float32/dims=2L	998083 ns	778875 ns	1.28
array/reductions/mapreduce/Int64/1d	585875 ns	583000 ns	1.00
array/reductions/mapreduce/Int64/dims=1	621792 ns	455542 ns	1.36
array/reductions/mapreduce/Int64/dims=1L	1035042 ns	1048250 ns	0.99
array/reductions/mapreduce/Int64/dims=2	782250 ns	516750 ns	1.51
array/reductions/mapreduce/Int64/dims=2L	2188333 ns	2170125 ns	1.01
array/reductions/reduce/Float32/1d	385375 ns	390083 ns	0.99
array/reductions/reduce/Float32/dims=1	280958 ns	323834 ns	0.87
array/reductions/reduce/Float32/dims=1L	613083 ns	619167 ns	0.99
array/reductions/reduce/Float32/dims=2	231959 ns	198166 ns	1.17
array/reductions/reduce/Float32/dims=2L	458667 ns	450084 ns	1.02
array/reductions/reduce/Int64/1d	615750 ns	582542 ns	1.06
array/reductions/reduce/Int64/dims=1	629041 ns	554334 ns	1.13
array/reductions/reduce/Int64/dims=1L	1031292 ns	1045250 ns	0.99
array/reductions/reduce/Int64/dims=2	258333 ns	241334 ns	1.07
array/reductions/reduce/Int64/dims=2L	647167 ns	636542 ns	1.02
array/shared/copy	136875 ns	143416 ns	0.95
array/shared/copyto!/cpu_to_gpu	37958 ns	38000 ns	1.00
array/shared/copyto!/gpu_to_cpu	38875 ns	38583 ns	1.01
array/shared/copyto!/gpu_to_gpu	38750 ns	38208 ns	1.01
array/shared/iteration/findall/bool	1076667 ns	1137958 ns	0.95
array/shared/iteration/findall/int	1251834 ns	1270792 ns	0.99
array/shared/iteration/findfirst/bool	958667 ns	640708 ns	1.50
array/shared/iteration/findfirst/int	962625 ns	812541 ns	1.18
array/shared/iteration/findmin/1d	1096875 ns	975458 ns	1.12
array/shared/iteration/findmin/2d	1070792 ns	1081334 ns	0.99
array/shared/iteration/logical	1559209 ns	1712625 ns	0.91
array/shared/iteration/scalar	4017.8571428571427 ns	4035.714285714286 ns	1.00
array/sorting/1d	1787667 ns	1853000 ns	0.96
array/sorting/2d	8282541 ns	8321959 ns	1.00
integration/byval/reference	1103708 ns	1099458 ns	1.00
integration/byval/slices=1	1117209 ns	1110709 ns	1.01
integration/byval/slices=2	2023917 ns	2040334 ns	0.99
integration/byval/slices=3	6636625 ns	15597292 ns	0.43
integration/metaldevrt	376500 ns	376500 ns	1
kernel/indexing	196208 ns	171208 ns	1.15
kernel/indexing_checked	384917 ns	308250 ns	1.25
kernel/launch	1895.8 ns	1916.7 ns	0.99
kernel/rand	386916 ns	314875 ns	1.23
latency/import	2040973208 ns	2068127667 ns	0.99
latency/precompile	38933808083 ns	39359634375 ns	0.99
latency/ttfp	2380935584 ns	2408685084 ns	0.99
metal/synchronization/context	595.2067039106146 ns	593.3988764044943 ns	1.00
metal/synchronization/stream	332.96 ns	331.6036866359447 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.

[christiangnrd]

I think this is ready