Fix implementation of the structure prior term of the BDeu score

[ZJsheep]

According to the Bernoulli edge-based structure prior combined with the BDeu marginal likelihood, the structure prior can be written as

$$ \log p(G) = \sum_{i=1}^{d} \left[ |PA_i| \log\left(\frac{\kappa}{d-1}\right) + (d - 1 - |PA_i|)\log\left(1 - \frac{\kappa}{d-1}\right) \right], $$

where $d$ denotes the number of variables (features), since each node can have at most $d-1$ possible parents.

In the original implementation, however, the Data array has shape (number_of_samples, number_of_variables). Therefore, Data.shape[1] should be used instead of Data.shape[0] when computing the structure prior.

[zhi-yi-huang]

Thanks for the fix!