Evaluating CxG Generalisation in LLMs via Construction-Based NLI Fine Tuning