TLoRA: Tri-Matrix Low-Rank Adaptation of Large Language Models