fix: remove incorrect start-vllm.sh that would break orchestrator architecture

start-vllm.sh

@@ -1,12 +0,0 @@
-#!/bin/bash
-# Start vLLM server with OpenAI-compatible API
-# This uses vLLM's built-in server which properly handles streaming
-
-vllm serve Qwen/Qwen2.5-7B-Instruct \
-  --host 0.0.0.0 \
-  --port 9000 \
-  --tensor-parallel-size 1 \
-  --gpu-memory-utilization 0.85 \
-  --max-model-len 4096 \
-  --download-dir /workspace/huggingface_cache \
-  --trust-remote-code