Exact Acceleration of Subgraph Graph Neural Networks by Eliminating Computation Redundancy