Multi-Granularity Prediction with Learnable Fusion for Scene Text Recognition