refactor server.py to use vLLM Python API and add permission config

server.py: switch from os.execvp(vllm) to make_arg_parser + asyncio.run(run_server(args)) for better error handling; add GPU selection via CLI argument
opencode.json: add default model field and permission.question=ask to prevent auto-triggering question tool

opencode.json

@@ -1,5 +1,6 @@
 {
   "$schema": "https://opencode.ai/config.json",
+  "model": "Agent/Qwen2.5-7B-Instruct",
   "provider": {
     "Agent": {
       "api": "openai",
@@ -17,6 +18,9 @@
       }
     }
   },
+  "permission": {
+    "question": "ask"
+  },
   "experimental": {
     "mcp_timeout": 60000
   }