Learning Trajectory-Aware Transformer for Video Super-Resolution