Which test statistic is the most appropriate for testing $H_0: \theta = \theta_0$ when the likelihood function is costly to maximize but the gradient of the log-likelihood is easy to compute at $\theta_0$?