Balanced and Elastic End-to-end Training of Dynamic LLMs