fix(openai): Use realistic Choice objects in manual output counting test

The test used MagicMock(message="one") where message was a plain string,
but the real OpenAI API returns Choice objects with message.content. The
counting code checks hasattr(choice.message, "content"), which failed on
strings, so manual token counting was never exercised. Use real Choice
and ChatCompletionMessage objects and fix the expected output_tokens.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: ericapisani

tests/integrations/openai/test_openai.py

@@ -2105,12 +2105,24 @@ def count_tokens(msg):
     response.usage.prompt_tokens = 20
     response.usage.total_tokens = 20
     response.choices = [
-        mock.MagicMock(message="one"),
-        mock.MagicMock(message="two"),
-        mock.MagicMock(message="three"),
+        Choice(
+            index=0,
+            finish_reason="stop",
+            message=ChatCompletionMessage(role="assistant", content="one"),
+        ),
+        Choice(
+            index=1,
+            finish_reason="stop",
+            message=ChatCompletionMessage(role="assistant", content="two"),
+        ),
+        Choice(
+            index=2,
+            finish_reason="stop",
+            message=ChatCompletionMessage(role="assistant", content="three"),
+        ),
     ]
     messages = []
-    streaming_message_responses = []
+    streaming_message_responses = None
 
     with mock.patch(
         "sentry_sdk.integrations.openai.record_token_usage"
@@ -2127,7 +2139,7 @@ def count_tokens(msg):
             span,
             input_tokens=20,
             input_tokens_cached=None,
-            output_tokens=None,
+            output_tokens=11,
             output_tokens_reasoning=None,
             total_tokens=20,
         )