Exploiting Block Coordinate Descent for Cost-Effective LLM Model Training