Scaling Vision-Language-Action Models with 100K+ Hours of Real-World Data