fix: Ensure stop_reason is captured by yielding on RawMessageDeltaEvent

The previous implementation had a timing issue where stop_reason arrived
after the last chunk was yielded. This fix ensures that when we receive
RawMessageDeltaEvent (which contains the stop_reason), we yield an
additional chunk with the updated metadata.

This addresses the test failure identified by @AstraBert where:
- Content chunks were yielded before stop_reason arrived
- stop_reason was None in the last chunk
- RawMessageDeltaEvent came after ContentBlockStopEvent

Now the flow is:
1. ContentBlockDeltaEvent -> yield chunks with content
2. ContentBlockStopEvent -> yield final content chunk
3. RawMessageDeltaEvent -> yield metadata update with stop_reason
4. Tests now pass with stop_reason correctly captured

Author: zhangzhefang-github

llama-index-integrations/llms/llama-index-llms-anthropic/llama_index/llms/anthropic/base.py

@@ -619,6 +619,21 @@ def gen() -> Generator[AnthropicChatResponse, None, None]:
                         }
                     if hasattr(r, "delta") and hasattr(r.delta, "stop_reason"):
                         stop_reason = r.delta.stop_reason
+
+                    # Yield a final chunk with updated metadata including stop_reason
+                    yield AnthropicChatResponse(
+                        message=ChatMessage(
+                            role=role,
+                            blocks=content,
+                            additional_kwargs={
+                                "usage": usage_metadata if usage_metadata else None,
+                                "stop_reason": stop_reason,
+                            },
+                        ),
+                        citations=cur_citations,
+                        delta="",
+                        raw=dict(r),
+                    )
                 elif isinstance(r, RawMessageStopEvent):
                     # Final event - no additional data to capture
                     pass
@@ -851,6 +866,21 @@ async def gen() -> ChatResponseAsyncGen:
                         }
                     if hasattr(r, "delta") and hasattr(r.delta, "stop_reason"):
                         stop_reason = r.delta.stop_reason
+
+                    # Yield a final chunk with updated metadata including stop_reason
+                    yield AnthropicChatResponse(
+                        message=ChatMessage(
+                            role=role,
+                            blocks=content,
+                            additional_kwargs={
+                                "usage": usage_metadata if usage_metadata else None,
+                                "stop_reason": stop_reason,
+                            },
+                        ),
+                        citations=cur_citations,
+                        delta="",
+                        raw=dict(r),
+                    )
                 elif isinstance(r, RawMessageStopEvent):
                     # Final event - no additional data to capture
                     pass