Scalable Training of Graph Foundation Models for Atomistic Materials Modeling: A Case Study with HydraGNN