fix: httpcore deadlock in CI by properly closing streaming responses (#4335)

# What does this PR do?

The test_conversation_error_handling test was timing out in CI with a
deadlock in httpcore's connection pool. The root cause was the preceding
test_conversation_multi_turn_and_streaming test, which broke out of the
streaming response iterator early without properly closing the
underlying HTTP connection.

When a streaming response iterator is abandoned mid-stream, the HTTP
connection remains in an incomplete state. Since the openai_client
fixture is session-scoped, subsequent tests reuse the same httpcore
connection pool. The dangling connection causes the pool's internal lock
to deadlock when the next test attempts to acquire a new connection.

The fix wraps the streaming response in a context manager, which ensures
the connection is properly closed when exiting the with block, even when
breaking out of the loop early. This is a best practice when working
with streaming HTTP responses that may not be fully consumed.

Signed-off-by: Sébastien Han <seb@redhat.com>

tests/integration/responses/test_conversation_responses.py

@@ -43,19 +43,18 @@ class TestConversationResponses:
             conversation=conversation.id,
         )
 
-        # Second turn with streaming
-        response_stream = openai_client.responses.create(
+        # Second turn with streaming - use context manager to ensure proper connection cleanup
+        with openai_client.responses.create(
             model=text_model_id,
             input=[{"role": "user", "content": "Say goodbye"}],
             conversation=conversation.id,
             stream=True,
-        )
-
-        final_response = None
-        for chunk in response_stream:
-            if chunk.type == "response.completed":
-                final_response = chunk.response
-                break
+        ) as response_stream:
+            final_response = None
+            for chunk in response_stream:
+                if chunk.type == "response.completed":
+                    final_response = chunk.response
+                    break
 
         assert response1.id != final_response.id
         assert len(response1.output_text.strip()) > 0
@@ -67,12 +66,7 @@ class TestConversationResponses:
 
     @pytest.mark.timeout(60, method="thread")
     def test_conversation_context_loading(self, openai_client, text_model_id):
-        """Test that conversation context is properly loaded for responses.
-
-        Note: 60s timeout added due to CI-specific deadlock in pytest/OpenAI client/httpx
-        after running 25+ tests. Hangs before first HTTP request is made. Works fine locally.
-        Investigation needed: connection pool exhaustion or event loop state issue.
-        """
+        """Test that conversation context is properly loaded for responses."""
         conversation = openai_client.conversations.create(
             items=[
                 {"type": "message", "role": "user", "content": "My name is Alice. I like to eat apples."},