INT4 quantization of HuggingFaceTB/SmolVLM-Instruct (LLM backbone quantized, vision encoder fp16)

initial commit