Declare Shape, Reshape, Transpose, Squeeze, Unsqueeze for opsets 21, 23 on CUDA

[xadupre]

Description

Fixes #26065.

[Copilot]

Pull Request Overview

This PR adds support for Shape, Reshape, Transpose, Squeeze, and Unsqueeze operations for ONNX opsets 21 and 23 on the CUDA execution provider. This addresses issue #26065 by declaring these tensor operations for the newer opset versions.

Key changes include:

• Adding versioned kernel declarations for opsets 21-22 and new opset 23 kernel declarations
• Updating existing opset 13 and 19 kernels to be versioned up to opset 20
• Adding comprehensive test coverage for the new opset versions

Reviewed Changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated no comments.

File	Description
onnxruntime/core/providers/cuda/tensor/*.cc	Add CUDA kernel declarations for Shape, Reshape, Transpose, Squeeze, Unsqueeze ops for opsets 21-23
onnxruntime/core/providers/cuda/cuda_execution_provider.cc	Update kernel class declarations and registrations to support versioned kernels and new opsets
onnxruntime/test/providers/cpu/tensor/*.cc	Add test cases for opsets 21 and 23
onnxruntime/test/optimizer/transpose_optimizer_test.cc	Update transpose optimizer tests to include opsets 21 and 23

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[yuslepukhin]

Does it fix it?

[xadupre]

It works fine on the model I used to test. I did not check the whole list of missing kernels declarations. We could also change the logic behind the kernel selection. An operator is often upgraded because it supports more types but onnxruntime does not always implements the kernel for the new types.