On Maximum Likelihood Training of Score-Based Generative Models