feat: Add native Gemma3 support for ONNX export

[Ankit-06679]

🎯 Problem

Users trying to export fine-tuned Gemma3 models to ONNX format encounter this error:

This prevents users from converting their Gemma3 models to ONNX format for deployment and optimization.

🔧 Solution

This PR adds native Gemma3 support to optimum by implementing the Gemma3OnnxConfig class and properly registering it in the tasks manager.

📝 Changes Made

✨ Core Implementation

Added Gemma3OnnxConfig class** in optimum/exporters/onnx/model_configs.py

• Inherits from TextDecoderOnnxConfig (same pattern as Gemma2)
• Uses GemmaDummyPastKeyValuesGenerator for compatibility
• Supports text generation and classification tasks
• Requires transformers >= 4.50.0 (when Gemma3 was introduced)

🔧 Supporting Changes

• Updated optimum/exporters/onnx/utils.py** - Added "gemma3" to MODEL_TYPES_REQUIRING_POSITION_IDS
• Updated test files** - Added Gemma3OnnxConfig imports and test model mappings
• Updated supported architectures** - Added proper version checks for Gemma3 support

🧪 Testing & Verification

• ✅ All syntax checks pass
• ✅ Configuration properly registered in tasks manager
• ✅ Follows existing Gemma/Gemma2 implementation patterns
• ✅ No breaking changes to existing functionality
• ✅ Comprehensive verification script confirms all changes

📋 Usage Example

After this change, users can export Gemma3 models seamlessly:

from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForCausalLM

# This will now work without errors! 🎉
model_id = "path/to/gemma3/model"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = ORTModelForCausalLM.from_pretrained(model_id, export=True)
model.save_pretrained('output_onnx/')

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[IlyasMoutawwakil]

Thanks ! This was already done in #70
What would be interesting is adding support for gemma3 VLM variant.

[Ankit-06679]

Thanks ! This was already done in #70 What would be interesting is adding support for gemma3 VLM variant.

Oh,I didn't notice that. I will try the gemma3 VLM variant if possible.

[IlyasMoutawwakil]

yes please do, you can use #50 for reference