Quantized-Tinyllava: a new multimodal foundation model enables efficient split learning