Exponential Regret Bounds for Gaussian Process Bandits with Deterministic Observations