Pareto-Optimal Learning from Preferences with Hidden Context