Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LLM Reasoning