Dissecting the Ledger: Locating and Suppressing "Liar Circuits" in Financial Large Language Models