Accelerating Generic Graph Neural Networks via Architecture, Compiler, Partition Method Co-Design