This removes an inconsistency between Numpy and Aesara broadcasting rules, where a variable dimension with unknown shape was always assumed to be non-broadcastable (i.e., different than 1)