Reversible Foundations: Training a 120B Sparse MoE through State-Preserving Scaling