Inverse articulated-body dynamics from video via variational sequential Monte-Carlo