Efficient human-in-the-loop MPC tuning with multi-task preferential Bayesian optimization