fix: implement conditional probability scoring

fix: shuffle MCQ candidates, fix length bias, add difficulty tracking

feat: adapt eval script for Kaggle environment and 19GB disk limits

fix: enforce GPU 0 device_map and enable CPU offload for 8-bit models to bypass conservative memory dispatcher

fix: use BitsAndBytesConfig for newer transformers and clear disk cache to prevent IO Error

fix: total_mem -> total_memory (correct PyTorch API)

feat: v4.0 crash-safe eval suite — checkpoint/resume, aggressive VRAM cleanup, 12 mainstream models

docs: rewrite paper to professional 2-column format with expanded analysis; feat: add Google Colab evaluation script