fix: properly handle async service cleanup in LocalBackend.close()

The close() method was calling service.close() without checking if it was
async, causing resource leaks when services had async cleanup methods.
This fix:
- Makes LocalBackend.close() async to match the base Backend class
- Properly awaits async service close methods
- Maintains backward compatibility for sync close methods
- Handles __exit__ context manager compatibility

Prevents GPU memory leaks and zombie processes in production deployments.

Author: prashant182

src/art/local/backend.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 import math
 from datetime import datetime
@@ -82,16 +83,25 @@ def __enter__(self):
         return self
 
     def __exit__(self, *excinfo):
-        self.close()
-
-    def close(self):
+        try:
+            asyncio.get_running_loop()
+            # If we're already in an async context, we can't await here
+            # The caller will need to handle cleanup manually
+        except RuntimeError:
+            # No event loop running, we can create one
+            asyncio.run(self.close())
+
+    async def close(self):
         """
         If running vLLM in a separate process, this will kill that process and close the communication threads.
         """
         for _, service in self._services.items():
             close_method = getattr(service, "close", None)
             if callable(close_method):
-                close_method()
+                if asyncio.iscoroutinefunction(close_method):
+                    await close_method()
+                else:
+                    close_method()
 
     async def register(
         self,