vMCU: Coordinated Memory Management and Kernel Optimization for DNN Inference on MCUs