Don't poll before blocking synchronization, and only poll once otherwise

[maleadt]

Since #3315, synchronize_object queried the stream or event before every synchronization, only to decide whether to collect garbage. In the blocking path, that query costs about as much as synchronizing an idle stream (~350 ns on the benchmark machine). In the spinning path, cooperative_wait then queried again. This explains the cuda/synchronization/stream/{blocking,nonblocking} regressions on the dashboard.

Now the object is polled once up front, and only when cooperative_wait would poll it anyway (nonblocking, spinning). If it's done, we return right away, and events still go through synchronize_completed. The GC pause and draining of retired resources happen only when we're about to wait. Blocking and spin=false synchronization don't poll at all, as before #3315.

Idle-GPU benchmarks (RTX 5080, Julia 1.12, 3 pinned cores, median of 3 alternating runs, ns):

benchmark	main	this stack	this PR
stream/auto	176	196	86
stream/blocking	194	214	195
stream/nonblocking	1977	2475	2187

The context/auto regression on the dashboard (~1 → ~6.8 µs) is intended and not addressed here. Before #3315, device_synchronize() polled the legacy stream, which doesn't cover CUDA.jl's non-blocking task streams, and then blocked the thread in cuCtxSynchronize. A context can't be polled, so it now always goes through a worker, and context/auto costs the same as context/nonblocking.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: a581505	Previous: ae637e5	Ratio
array/accumulate/Float32/1d	105613 ns	105001 ns	1.01
array/accumulate/Float32/dims=1	77837 ns	77262 ns	1.01
array/accumulate/Float32/dims=1L	1675150 ns	1684940 ns	0.99
array/accumulate/Float32/dims=2	148391 ns	148043 ns	1.00
array/accumulate/Float32/dims=2L	685943 ns	684338 ns	1.00
array/accumulate/Int64/1d	124490 ns	123503 ns	1.01
array/accumulate/Int64/dims=1	82609 ns	81902 ns	1.01
array/accumulate/Int64/dims=1L	1805589 ns	1803636 ns	1.00
array/accumulate/Int64/dims=2	163785 ns	163105 ns	1.00
array/accumulate/Int64/dims=2L	989761 ns	988275 ns	1.00
array/broadcast	17392 ns	16981 ns	1.02
array/broadcast launch	7833 ns	7497.25 ns	1.04
array/construct	935.1538461538462 ns	886.76 ns	1.05
array/copy	17390 ns	16325 ns	1.07
array/copyto!/cpu_to_gpu	208722 ns	206249 ns	1.01
array/copyto!/gpu_to_cpu	241867 ns	240560 ns	1.01
array/copyto!/gpu_to_gpu	9054 ns	8967 ns	1.01
array/iteration/findall/bool	140114 ns	135815 ns	1.03
array/iteration/findall/int	150507 ns	145400 ns	1.04
array/iteration/findfirst/bool	75376 ns	71077 ns	1.06
array/iteration/findfirst/int	76921 ns	73309 ns	1.05
array/iteration/findmin/1d	67992 ns	59845 ns	1.14
array/iteration/findmin/2d	103964 ns	103127 ns	1.01
array/iteration/logical	201008 ns	187405 ns	1.07
array/iteration/scalar	53566 ns	59115 ns	0.91
array/permutedims/2d	50102 ns	49364 ns	1.01
array/permutedims/3d	51685 ns	50795 ns	1.02
array/permutedims/4d	52799 ns	52271 ns	1.01
array/random/rand/Float32	11539 ns	11109 ns	1.04
array/random/rand/Int64	18823 ns	18411 ns	1.02
array/random/rand!/Float32	8344.333333333334 ns	8183 ns	1.02
array/random/rand!/Int64	16281 ns	15765 ns	1.03
array/random/randn/Float32	34895 ns	34501 ns	1.01
array/random/randn!/Float32	25010 ns	24403 ns	1.02
array/reductions/mapreduce/Float32/1d	32628 ns	30573 ns	1.07
array/reductions/mapreduce/Float32/dims=1	40334 ns	40213 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	53884 ns	53356 ns	1.01
array/reductions/mapreduce/Float32/dims=2	59585 ns	59358 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	72115 ns	71210 ns	1.01
array/reductions/mapreduce/Int64/1d	39669 ns	37469 ns	1.06
array/reductions/mapreduce/Int64/dims=1	43634 ns	42915 ns	1.02
array/reductions/mapreduce/Int64/dims=1L	89742 ns	89249 ns	1.01
array/reductions/mapreduce/Int64/dims=2	60996 ns	61370 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	88019 ns	87543 ns	1.01
array/reductions/reduce/Float32/1d	32847 ns	30615 ns	1.07
array/reductions/reduce/Float32/dims=1	40543 ns	40264 ns	1.01
array/reductions/reduce/Float32/dims=1L	53405 ns	52907 ns	1.01
array/reductions/reduce/Float32/dims=2	59613 ns	59462 ns	1.00
array/reductions/reduce/Float32/dims=2L	72568 ns	71728 ns	1.01
array/reductions/reduce/Int64/1d	39677 ns	37570 ns	1.06
array/reductions/reduce/Int64/dims=1	43568 ns	42862 ns	1.02
array/reductions/reduce/Int64/dims=1L	89505 ns	89319 ns	1.00
array/reductions/reduce/Int64/dims=2	60917 ns	61437 ns	0.99
array/reductions/reduce/Int64/dims=2L	88048 ns	86880 ns	1.01
array/reverse/1d	16561 ns	16128 ns	1.03
array/reverse/1dL	70409 ns	70115 ns	1.00
array/reverse/1dL_inplace	67824 ns	67602 ns	1.00
array/reverse/1d_inplace	9570 ns	9293 ns	1.03
array/reverse/2d	21350 ns	21118 ns	1.01
array/reverse/2dL	75395 ns	74528 ns	1.01
array/reverse/2dL_inplace	67896 ns	67778 ns	1.00
array/reverse/2d_inplace	10931 ns	10676 ns	1.02
array/sorting/1d	2850525 ns	2846981 ns	1.00
array/sorting/2d	1090831 ns	1088944 ns	1.00
array/sorting/by	3403924 ns	3400388 ns	1.00
cuda/graph/capture	42432 ns
cuda/graph/eager	41418 ns
cuda/graph/launch	14446 ns
cuda/synchronization/context/auto	7947 ns	6766.6 ns	1.17
cuda/synchronization/context/blocking	891.8205128205128 ns	793.7692307692307 ns	1.12
cuda/synchronization/context/nonblocking	7918 ns	6805.4 ns	1.16
cuda/synchronization/stream/auto	353.60648148148147 ns	691.0939597315436 ns	0.51
cuda/synchronization/stream/blocking	817.16 ns	868.875 ns	0.94
cuda/synchronization/stream/nonblocking	7426 ns	7338.5 ns	1.01
integration/byval/reference	148238 ns	147882 ns	1.00
integration/byval/slices=1	149280 ns	149033 ns	1.00
integration/byval/slices=2	291296 ns	291277 ns	1.00
integration/byval/slices=3	433701 ns	434160 ns	1.00
integration/cudadevrt	105543 ns	105116 ns	1.00
integration/volumerhs	9208712 ns	9206594 ns	1.00
kernel/indexing	11831 ns	11855 ns	1.00
kernel/indexing_checked	12393 ns	12209 ns	1.02
kernel/launch	2562.222222222222 ns	2433 ns	1.05
kernel/occupancy	959.95 ns	1008.8235294117648 ns	0.95
kernel/rand	15438 ns	14899 ns	1.04
latency/import	4316478842 ns	4305717656 ns	1.00
latency/precompile	5098977588 ns	5095420673 ns	1.00
latency/ttfp	4796690923 ns	4793974717 ns	1.00

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.10%. Comparing base (38d4a78) to head (a581505).

Additional details and impacted files

@@                  Coverage Diff                   @@
##           tb/event-ownership    #3360      +/-   ##
======================================================
+ Coverage               84.86%   85.10%   +0.23%     
======================================================
  Files                     194      194              
  Lines                   19960    19962       +2     
======================================================
+ Hits                    16939    16988      +49     
+ Misses                   3021     2974      -47     

☔ 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.