Optimizing Split Learning Latency in TinyML-Based IoT Systems