Lora fine-tuning of small language models for turkish contextual question answering and inference via an english reasoning–turkish answer strategy