Real-time Neural Network Inference on Extremely Weak Devices: Agile Offloading with Explainable AI