Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs