Where Should Optimizer State Live? Tiered State Allocation for Memory-Efficient Mixture-of-Experts Training