Reflecting on the State of Rehearsal-free Continual Learning with Pretrained Models