Training High-Performance Low-Latency Spiking Neural Networks by Differentiation on Spike Representation