Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification