Towards Improving Harmonic Sensitivity and Prediction Stability for Singing Melody Extraction