Fix LLMTask._finish_function_call to bypass deferral

self.queue_frame would defer the LLMMessagesAppendFrame because
_finish_function_call always runs inside a tool call. The subsequent
_flush_pipeline() then returned before the goodbye/handoff LLM output
was actually delivered. Use super().queue_frame to push the frame
straight into the pipeline, matching the pattern used in
_flush_pipeline().

src/pipecat/tasks/llm/llm_task.py

@@ -327,7 +327,10 @@ class LLMTask(PipelineTask):
             messages: Optional LLM messages to inject before completing.
         """
         if messages:
-            await self.queue_frame(LLMMessagesAppendFrame(messages=messages, run_llm=True))
+            # Bypass our deferral override: this runs inside a tool call, so
+            # self.queue_frame would defer the frame and the flush below would
+            # return before the LLM output is delivered.
+            await super().queue_frame(LLMMessagesAppendFrame(messages=messages, run_llm=True))
             await self._flush_pipeline()
 
         if not result_callback: