Scaling Large Language Model Training on Frontier with Low-Bandwidth Partitioning