when ARUCO tracker fails, observation should use last tracked position instead of position from last step

[Stapelzeiger]

caching the last valid position in step adds much more latency than needed. We should add a cache of last valid position updated at every frame.

[Stapelzeiger]

Currently this results in exactly 0 reward for walking tasks, which we use to troubleshoot bad tracking. We would have to add some other stats to monitor tracking quality ater this gets fixed.