Add Gemma-3 vision tutorial to ONNX Runtime GenAI

[kunal-vaishnavi]

Description

This PR adds a tutorial to show how to create the ONNX models for Gemma-3 vision (4B, 12B, 27B).

Motivation and Context

This PR requires the changes from the following PRs.

• Support for gemma3 model #1374
• Fix accessing final norm for Gemma-3 models #1687
• Fix runtime bugs with multi-modal models #1701
• Enable bfloat16 multi-modal models #1786

This PR also resolves the following issues.

• support gemma 3 #1329
• OGA failing to run Gemma-3-4b OGA model. #1536
• Unable to run Gemma-3-4b-it ONNX Model. #1655
• bad_allocation error when providing image input to Gemma-3-4b-it CPU FP32 ONNX model #1698

[snnn]

We have tutorial documentations in the main ORT repo and here. How are they organized?

[kunal-vaishnavi]

We have tutorial documentations in the main ORT repo and here. How are they organized?

The tutorial organization is a relic of the team's historical work. The ORT transformers folder was the original location for any tutorials. Once ORT GenAI was created, any tutorials specific to running multi-modal models were stored in the examples folder here. All of the ORT GenAI tutorials require modeling code changes and currently those modified files are stored in the same Hugging Face repos that contain the optimized and quantized ONNX models.

Now that we have an Olive recipes repo for getting models, we can move the remaining tutorials there. However, we need to carefully do it in migration phases as we have production customers who depend on the existing tutorials and their associated modeling code changes.