On the Relation between Internal Language Model and Sequence Discriminative Training for Neural Transducers