Intra-request branch orchestration for efficient LLM reasoning