Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models