On the relation between trainability and dequantization of variational quantum learning models