GOAT-TTS: Expressive and Realistic Speech Generation via A Dual-Branch LLM