Robust Multi-Objective Controlled Decoding of Large Language Models