Add a simulation step in the reset.

[idigitopia]

We are using rigid_body_states to update the observation. however these are not updated till a simulation step happens. so technically we are getting wrong observations at every reset.

[xbpeng]

Unfortunately this will cause other issues. It will result in the environments being advanced two timesteps instead of one whenever one of the envs terminate, and only advancing one timestep when no envs terminate. We can't have that kind of randomness in the dynamics.