Conformal Arbitrage: Risk-Controlled Balancing of Competing Objectives in Language Models