Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks