Dynamic Subset Tuning: Expanding the Operational Range of Parameter-Efficient Training for Large Language Models