[CPU] Handle ONNX domain Gelu and HardSigmoid activations in the NCHWc transformer suite

[hariharans29]

Description

As title

Motivation and Context

Tiny continuation to #27691

[Copilot]

Pull request overview

Extends the CPU NCHWc transformer pipeline to treat HardSigmoid like the other supported elementwise activations, enabling both (a) reorder-free passthrough in NCHWc subgraphs and (b) fusion into an upstream NCHWc Conv when eligible.

Changes:

• Add HardSigmoid handling in NchwcTransformerImpl::TransformActivation, including passing alpha/beta via activation_params when fusing into NCHWc Conv.
• Recognize HardSigmoid nodes in the transformer dispatch list (opset since versions 6 and 22).
• Add/extend optimizer tests to cover both “no extra reorder” behavior and single-consumer Conv+Activation fusion (including alpha/beta attribute propagation).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/test/optimizer/nchwc_optimizer_test.cc	Adds HardSigmoid coverage for reorder-avoidance and Conv+activation fusion (including activation_params validation).
onnxruntime/core/optimizer/nchwc_transformer.cc	Enables HardSigmoid as a supported activation and fuses it into NCHWc Conv with correct default/custom alpha/beta handling.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

You can commit the suggested changes from lintrunner.

[tianleiwu]

Optimizer Tests (onnxruntime/test/optimizer/nchwc_optimizer_test.cc)

Concern:

• ⚠️ Missing HardSigmoid in the no-fusion guard test (Nitpick): The new ActivationSingleConsumerConvNoFusion test only covers Gelu (both ONNX and MS-domain) and QuickGelu — activations that should never fuse. This is intentional and correct. However, there's no test for a scenario where HardSigmoid is present but the Conv already has an activation fused (double-fusion prevention). That edge case is covered by the graph_utils::GetNodeAttribute(nchwc_node, "activation") == nullptr guard in the transformer, but there's no explicit test for it. This is a pre-existing gap, not introduced by this PR.

Verdict

APPROVE — Well-structured, minimal change with correct defaults, proper kernel-level support already in place, and thorough test coverage. No functional issues found.

[tianleiwu]

Please run lintrunner