Gaussian process based nonlinear latent structure discovery in multivariate spike train data