Towards a Holistic Framework for Multimodal Large Language Models in Three-dimensional Brain CT Report Generation