Effective Text Adaptation for LLM-based ASR through Soft Prompt Fine-Tuning