Optimistic Thompson Sampling for No-Regret Learning in Unknown Games