Improve VLM client lifecycle, retries, metrics, and cache keys

- Adds abstract aclose() and shared load_prompt() helper to BaseVLMClient.

- OllamaClient now logs latency, tokens, and prompt hash per call.

- Retries cover HTTP 429 and 5xx in addition to network/timeout errors.

- Cache key now includes response_format and max_tokens.

- Warns when JSON schema generation falls back to plain 'json' format.

- Adds FakeVLMClient.aclose() and cache unit tests.

tests/conftest.py

@@ -52,3 +52,7 @@ class FakeVLMClient(BaseVLMClient):
             if key in system_prompt or key in user_prompt:
                 return VLMResponse(content=json.dumps(value), model="fake")
         return VLMResponse(content=json.dumps({}), model="fake")
+
+    async def aclose(self) -> None:
+        """The fake client holds no resources."""
+        return None