Quantifying Training Difficulty and Accelerating Convergence in Neural Network-Based PDE Solvers