Fine-Tuning Qwen for Image-to-Text Extraction on a Single T4 GPU Using Unsloth and TRL