Generalized Linear Bandits: Almost Optimal Regret with One-Pass Update