Selective Adapter Freezing for Memory-Efficient Fine-Tuning of Language Models