The onnxruntime-genai.qnn encounters hallucination issues when using Phi-4-mini-instruct

**Describe the bug**
The onnxruntime-genai.qnn encounters hallucination issues when using Phi-4-mini-instruct

**To Reproduce**
Steps to reproduce the behavior:
1. Use VSCode to convert the Microsoft/Hi-4-mini-instruct model in Huggingface.
2. Use C #&OnnxRuntime. Gneai.qnn to call the model.
3. Send "如何制作一份煎牛排"
4. The following situation will occur

**Expected behavior**
Normal output of the process of making steak, even if there is no output of the process of making steak, at least the text content should be output normally.

**Screenshots**

**Desktop (please complete the following information):**
- OS:Windows 11 Preview
- Version [e.g. 22]
- Microsoft.ML.OnnxRuntimeGenAI.QNN[0.8.3]

**Additional context**
[Information related to model transformation and reasoning problem content](https://github.com/microsoft/Olive/issues/1962#issuecomment-3101199494)

Start by reproducing the Chinese steak-making prompt with Microsoft/Hi-4-mini-instruct, Microsoft.ML.OnnxRuntimeGenAI.QNN 0.8.3, and the C# caller on Windows 11. Review the linked Olive issue comment for the model-conversion and reasoning context, then determine whether conversion or QNN inference causes the hallucinated output. Done means normal text output is restored or the failure is isolated and documented.

Written by the indexing model from the issue text.