upstream connect error or disconnect/reset before headers. reset reason: remote connection failure, transport failure reason: delayed connect error: Connection refuse

### 🐛 Describe the bug

```
root@benchmark-client:/vllm-workspace/benchmarks# curl http://10.249.253.10:80/v1/chat/completions -H "Content-Type: application/json" -H "routing-strategy: pd" -d '{
"model": "qwen3-8B",
"messages": [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "help me write a random generator in python"}
]
}'
upstream connect error or disconnect/reset before headers. reset reason: remote connection failure, transport failure reason: delayed connect error: Connection refused

```

### Steps to Reproduce

1p1d vllm deployment
model is loading, send request before it's ready

### Expected behavior

should return json response

### Environment

nightly

Start by reproducing the request against the 1P1D vLLM deployment while the model is loading, then inspect the deployment and upstream connection behavior. Done means the same chat-completions request returns a JSON response instead of the connection-refused error once the model is ready.

Written by the indexing model from the issue text.