On the Role of Entropy-based Loss for Learning Causal Structures with Continuous Optimization