Hey!
There is no matrix normal distribution in Stan, although you could build it yourself. However, if you want to give the entries in B independent normal distributions, you could just write to_vector(B) ~ normal(...,...) as your prior. You gain nothing using something like B[1,] ~ multi_normal(mu_u,diag_matrix(sigma_u)), because the diagonal matrix implies they are the entries in B are not correlated. Using univariate normals thus would speed up and simplify your model.
You declared log_y as transformed parameter, but you really should declare it as transformed data. If you put a prior on a transformed parameter Stan thinks you might have a change of variable and need to apply a Jacobian correction. But in your case you just transform the data, so the warning should go away when log_y is declared in the transformed data block.
Cheers!
Max