fix: decode slice (gen_ids[0,-max_new:]) + defaults temp=0.2/max=2000/rep=1.05

fix: BF16 dtype for PULSE-7B (mat1/mat2 dtype mismatch)

fix: pass input_ids as inputs= (LLaVA generate signature)

fix: force llava prefix in model_name so builder loads vision tower

Fix: unset HF_MODEL_DIR so toolkit does not conflict with HF_MODEL_ID

Vendor LLaVA v1.2.0 + thin handler (no pip install at runtime)