An Experimental Study of Split-Learning TinyML on Ultra-Low-Power Edge/IoT Nodes