CE-LoRA: Computation-Efficient LoRA Fine-Tuning for Language Models