On the Sample Complexity of Learning Sum-Product Networks