Generalized Bayesian Upper Confidence Bound with Approximate Inference for Bandit Problems