Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing