implement zero offset

[codekansas]

proper zero offset implementation, where the observation offset matches the engine offset

for reading qpos: if current "true" qpos is 0 but the zero position is set to 0.1, then on the real robot, we will read qpos as -0.1. similarly, if the policy outputs a target position of 0, then on the real robot, we will actually go to the target position of 0.1