Profit over Proxies: A Scalable Bayesian Decision Framework for Optimizing Multi-Variant Online Experiments