How Much Do Large Language Models Know about Human Motion? A Case Study in 3D Avatar Control