Promises and Pitfalls of the Linearized Laplace in Bayesian Optimization