fix: use generate_training_narrative + eval sampling params (temp=0.1, top_p=0.9)

feat: local Titan (base model, adapter off) - src/demo/model_loader.py

fix: two-step adapter loading + Nemotron v2 (4/10) - src/demo/model_loader.py

fix: center layout, dark body bg, model switching eviction - src/demo/model_loader.py

feat: multi-model support (Gemma + Nemotron) - src/demo/model_loader.py

fix: update src/demo/model_loader.py

Upload src/demo/model_loader.py with huggingface_hub