Support Gemma3 with Clip fused attention

[titaiwangms]

Description

Essentially, the vision model is traced differently (this time it's without mask.), and the input indices of op.Add and op.MatMul can be different. Also, fp16 and fp32 need different tracing patterns (op.Cast).

1. Add another traced pattern to CLIP attention to cover no attention_mask case
2. Accept different index of input on op.Add and op.MatMul (be more general)
3. fp16 and fp32 shows different pattern (op.Cast after op.Softmax)
4. Refactor test_fastgelu.py to cover torch.onnx.export(..., dynamo=True)
5. Add gemma3 vision attention (SigLip) test to cover both fp16 and fp32

Motivation and Context

To optimize Gemma3 multi-modal model, the changes are needed. https://huggingface.co/google/gemma-3-4b-it

NOTE: some related follow-ups (upstream optimizations to onnxscript-optimizer):
microsoft/onnxscript#2158
microsoft/onnxscript#2156

[Copilot]

Pull Request Overview

This PR refactors and extends model tracing and fusion tests for the Gemma3 vision model while adding support for fp16 and fp32 tracing patterns and generalizing input indices for op.Add and op.MatMul. Key changes include:

• Introducing a new traced pattern for CLIP attention without an attention mask.
• Generalizing the input indices for op.Add and op.MatMul and differentiating tracing for fp16 and fp32.
• Refactoring test files to support dynamo export and adding new tests for Gemma3 vision attention (SigLip).

Reviewed Changes

Copilot reviewed 6 out of 8 changed files in this pull request and generated no comments.

File	Description
onnxruntime/test/python/transformers/test_gemma3_vision.py	Adds tests and model definitions for Gemma3 vision attention and layer normalization.
onnxruntime/test/python/transformers/test_gelu_fusions.py	Refactors Gelu fusion tests by parameterizing the tests for dynamo export.
onnxruntime/python/tools/transformers/fusion_fastgelu.py	Improves robustness in FastGelu fusion by handling cases where the root input comes directly from the graph input.
onnxruntime/python/tools/transformers/fusion_attention_clip.py	Generalizes pattern matching for attention fusion to support different index configurations and tensor formats.

Files not reviewed (2)

• tools/ci_build/github/linux/python/requirements.txt: Language not supported
• tools/ci_build/github/windows/python/requirements.txt: Language not supported

Comments suppressed due to low confidence (2)

onnxruntime/python/tools/transformers/fusion_attention_clip.py:152

• [nitpick] Consider using a named constant or adding a comment to clarify the purpose of 'None' as a wildcard in the index array for pattern matching.

                [1, None, 0, 0, 0],

onnxruntime/python/tools/transformers/fusion_attention_clip.py:232

• [nitpick] Consistently document or use a named constant for wildcard indices (such as 'None') to improve code clarity in pattern matching.

            q_nodes = self.model.match_parent_path(

[snnn]

Then changes under tools/ci_build are fine.