Currently only reverse-mode autodiff is supported for RB-PG where a rrule is implemented for the gradient of the Kalman filter loglikelihood.
Forward-mode diff is made challenging by the use of PDMats and Static Arrays.
Low priority since can just use reverse-mode for now.