[Feature Request] Extend CUDA ONNX Ops to latest opset version

[tianleiwu]

Describe the feature request

Summary

• Strict registry-level audit found 47 ONNX-domain CUDA ops whose highest registered opset is below the latest ONNX changelog opset.
• Priority breakdown: P1 28, P2 11, P3 8.
• Method: compare the highest opset registered in onnxruntime/core/providers/cuda/cuda_execution_provider.cc against the latest version entry in build/cuda/Debug/_deps/onnx-src/docs/Changelog.md.
• This report is intentionally strict about registration coverage. It does not assume a newer schema is automatically backward-compatible with an older kernel registration.

Priority Legend

• P1: foundational graph/model-import gaps that are more likely to block modern models or common exporter output.
• P2: secondary math/activation gaps that are real registration misses but are typically less central.
• P3: deprecated, generator, or otherwise lower-urgency gaps.

Gaps

Priority	Operator	CUDA Max	ONNX Latest	Gap	Registration	Changelog	PR
P1	GlobalAveragePool	1	22	21	cc:1843	md:26711	#27733
P1	GlobalMaxPool	1	22	21	cc:1852	md:26780	#27733
P1	ConstantOfShape	9	25	16	cc:1963	md:31634	#27728
P1	TopK	11	24	13	cc:2068	md:31274	#27735
P1	Flatten	13	25	12	cc:2334	md:31731	#27728
P1	RoiAlign	10	22	12	cc:2025	md:27917	#27646
P1	Size	13	25	12	cc:2263	md:32532	#27728
P1	ConvTranspose	11	22	11	cc:2080	md:26246	#27710
P1	MaxPool	12	22	10	cc:2121	md:27215	#27715
P1	Dropout	13	22	9	cc:2326	md:26459	Skip since it is training ops
P1	GRU	14	22	8	cc:2437	md:26601	#27738
P1	LSTM	14	22	8	cc:2440	md:26988	#27737
P1	RNN	14	22	8	cc:2434	md:27625	#27743
P1	Pad	18	25	7	cc:2522	md:32016	#27708
P1	Identity	19	25	6	cc:2567	md:31769	#27728
P1	If	19	25	6	cc:2568	md:31798	#27728
P1	Loop	19	25	6	cc:2569	md:31838	#27728
P1	Scan	19	25	6	cc:2590	md:32289	#27728
P1	DequantizeLinear	21	25	4	cc:2603	md:31672	#27745
P1	QuantizeLinear	21	25	4	cc:2617	md:32159	#27741
P1	Cast	23	25	2	cc:2686	md:31420	#27744
P1	Reshape	23	25	2	cc:2707	md:32239	#27742
P1	Shape	23	25	2	cc:2706	md:32455	#27734
P1	Squeeze	23	25	2	cc:2709	md:32563	#27739
P1	Transpose	23	25	2	cc:2708	md:32597	#27740
P1	Unsqueeze	23	25	2	cc:2710	md:32642	#27739
P2	Elu	6	22	16	cc:1677	md:26516	#27753
P2	InstanceNormalization	6	22	16	cc:1937	md:26943	#27758
P2	Selu	6	22	16	cc:1689	md:28021	#27753
P2	Cos	7	22	15	cc:1656	md:26313	#27756
P2	Sin	7	22	15	cc:1659	md:28062	#27756
P2	EyeLike	9	22	13	cc:1985	md:26555	#27757
P2	ThresholdedRelu	10	22	12	cc:2029	md:28209	#27753
P2	Round	11	22	11	cc:2106	md:27979	#27754
P2	Equal	13	19	6	cc:2220	md:22998	#27754
P2	ReduceMax	18	20	2	cc:2515	md:24331	#27755
P2	ReduceMin	18	20	2	cc:2508	md:24380	#27755
P3	RandomNormal	1	22	21	cc:2005	md:27726
P3	RandomNormalLike	1	22	21	cc:2006	md:27772
P3	RandomUniform	1	22	21	cc:2007	md:27822
P3	RandomUniformLike	1	22	21	cc:2008	md:27867
P3	Softplus	1	22	21	cc:1701	md:28120
P3	Softsign	1	22	21	cc:1695	md:28151
P3	Scatter	10	11	1	cc:1986	md:12972
P3	Upsample	9	10	1	cc:1957	md:10297

Notes

• Scatter and Upsample are deprecated ONNX operators, so they stay in the report but are ranked lower.
• Several P1 items are only 1-2 opsets behind, but they remain high priority because they are common graph-building or deployment operators.
• The earlier one-item Upsample note was too narrow; this wider pass covers the full ONNX-domain CUDA registration surface in the execution provider.

Describe scenario use case

To run model exported with lastest onnx opset version

[tianleiwu]

TODO:

• Add test opset-20 ReduceMax for int8_t and uint8_t