Armada: Memory-Efficient Distributed Training of Large-Scale Graph Neural Networks