Efficient non-conjugate Gaussian process factor models for spike count data using polynomial approximations