Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs