Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence