Designing losses for data-free training of normalizing flows on Boltzmann distributions