Model Merging is Secretly Certifiable: Non-Vacuous Generalisation Bounds for Low-Shot Learning