fix: use LiteLLM vLLM pass-through for qwen model

- Changed model from openai/qwen-2.5-7b to hosted_vllm/qwen-2.5-7b
- Implements proper vLLM integration per LiteLLM docs
- Fixes streaming response forwarding issue

ai/litellm-config.yaml

@@ -32,7 +32,7 @@ model_list:
   # Text Generation
   - model_name: qwen-2.5-7b
     litellm_params:
-      model: openai/qwen-2.5-7b
+      model: hosted_vllm/qwen-2.5-7b
       api_base: http://100.121.199.88:9000/v1  # RunPod GPU via Tailscale
       api_key: dummy
       rpm: 1000