So the estimation and control loops are usually triggered by incoming gyro samples. So, you only really slow down the whole system if the gyro input comes late.
If you want to add a delay to the outputs, you would have to add additional usleeps in the pwm_output_sim. However, this would probably only work with the current architecture, once lockstep support is merged, because at that point the simulator just waits for px4 to output the control signal.
If you want to simulate that delay, it might be an idea to just create a FIFO buffer of actuator outputs in jMAVSim/Gazebo, and therefore simulate the control signals taking some time to reach the actual actuators.