potential missing matrix transpose operation

[cynricfu]

According to my understanding of the GQE model, the relation matrix in the project function of BilinearMetapathDecoder should be transposed before performing the matrix multiplication: https://github.com/williamleif/graphqembed/blob/2775a242cf8a5df520530f2dd08aa8790f109ebf/netquery/decoders.py#L149-L150 But after I change it to self.mats[rel].t().mm(embeds), the model performance scores do not change much. I don't know which one is correct, the original code or my revision