History-Conditioned Spatio-Temporal Visual Token Pruning for Efficient Vision-Language Navigation