Scaling the Poisson GLM to massive neural datasets through polynomial approximations