ALoRA: Allocating Low-Rank Adaptation for Fine-tuning Large Language Models