Continual Gradient Low-Rank Projection Fine-Tuning for LLMs