Update deprecated CCCL API

[mc-nv]

Description

Update API

Motivation and Context

Address issues for: #24774

[mc-nv]

@microsoft-github-policy-service agree

[snnn]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

Has build failures

[snnn]

@mc-nv , it has some build failures. Please fix them.

[tianleiwu]

CI pipeline is using CUDA 12.2, which does not have ::cuda::maximum(). Need code like the following to fix build error:

onnxruntime/onnxruntime/contrib_ops/cuda/bert/embed_layer_norm_impl.cu

Lines 42 to 47 in b05b252

	#if CUDA_VERSION >= 12090
	::cuda::minimum<int> min;
	#else
	// Deprecated on CUDA 12.9
	cub::Min min;
	#endif

[mc-nv]

@snnn where I can find build environment configuration, or may be you can provide the version of CCCL installed.
At glance build uses CUDA 12.2 which 18 month old for NVIDIA.
Will it be possible to validate changes within 12.9 build environment?

Why we asking for it because CCCL API deprecation was introduced in version 2.8.0
https://github.com/NVIDIA/cccl/releases/tag/v2.8.0
And that version is 4 month old.

[snnn]

People are even asking for CUDA 11.x support. #25182

I think we still need to make the source code be compatible with CUDA 12.2 before the next release is done. Then we can discuss upgrading.

[mc-nv]

got it I will address those changes and get back to this PR

[tianleiwu]

/azp run Linux QNN CI Pipeline,Win_TRT_Minimal_CUDA_Test_CI,Windows ARM64 QNN CI Pipeline,Windows GPU Doc Gen CI Pipeline,Windows x64 QNN CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 5 pipeline(s).

[mc-nv]

I'm afraid I need to push one more commit towards to this change...

[snnn]

That's fine. Welcome to submit a new PR.

[tianleiwu]

I tested Windows build using cuda 12.9 with/without this PR. Both are good.

CUDA_VERSION is defined in cuda.h; CUDART_VERSION is defined in cuda_runtime_api.h. It is better to include cuda.h or cuda_runtime_api.h explicitly.