A Convex Relaxation Approach to Bayesian Regret Minimization in Offline Bandits