Direct Training Needs Regularisation: Anytime Optimal Inference Spiking Neural Network