Interpretable deep learning method integrating spatial self-attention for generating bias-corrected high-resolution GFS precipitation forecasts