Implement manual generator driving using next()/StopIteration to capture the return value (trailing usage dict) from inner stream_chat() implementations, enabling token tracking for streaming LLM calls. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>