Learning Fine-Grained Visual Understanding for Video Question Answering via Decoupling Spatial-Temporal Modeling