MEFT: Memory-Efficient Fine-Tuning through Sparse Adapter