Ground Slow, Move Fast: A Dual-System Foundation Model for Generalizable Vision-and-Language Navigation