Fix/honor user model name

[yuyongtan9-eng]

No description provided.

[ZebangCheng]

This PR correctly fixes the issue where the USER_MODEL_NAME environment variable was ignored: custom values are now passed to the OpenAI request as expected. Locally, the two new tests and the full 25-test suite pass, and all Python files compile successfully. However, I do not recommend merging it as-is. run_all_scenarios.sh currently defaults USER_MODEL_NAME to Qwen3.5-397B-A17B, while both the existing Python default and the canonical SiliconFlow model ID are Qwen/Qwen3.5-397B-A17B. Previously, the shell-script value had no effect because it was overwritten by the hard-coded Python configuration; after this PR, it will take effect for the first time. As a result, running bash run_all_scenarios.sh without an explicit override will change existing behavior and may cause all simulated-user requests to fail because of the invalid model ID. Please update the script default to : "${USER_MODEL_NAME:=Qwen/Qwen3.5-397B-A17B}", add USER_MODEL_NAME to .env.example, and include a regression test covering the batch entry point’s default value. Once these changes are made, the PR should be ready to merge.