Learning cardiac activation maps from 12-lead ECG with multi-fidelity Bayesian optimization on manifolds