Reconstructing scene from qpos and qvel

I've mentioned this briefly in #157, but I believe this deserves a separate thread.

In certain applications like Go-Explore (https://www.nature.com/articles/s41586-020-03157-9), it's critical for the user to be able to reset the environment to the state corresponding to the gathered observation, i.e., qpos and qvel.

In my experiments with Brax v2, I haven't been able to successfully reconstruct the simulator state from qpos and qvel alone--I suspect the 2nd order information is also necessary. According to MuJoCo docs:

> class mujoco_py.MjSimState
>
> Represents a snapshot of the simulator’s state.
>
> This includes time, qpos, qvel, act, and udd_state.
>
> Attributes
>
> Methods

Ignoring `time` and `udd_state`, `act` needs to be taken into account. Not sure by what mechanism MuJoCo does this, but we can gather that `qpos`, `qvel`, and `act` should be necessary and sufficient for reconstructing the state.

Current workaround for this is to actually store the states (i.e. in a batched instance of `State` class) gathered during exploration and to use that as the argument for `step()`. But this comes at a huge memory overhead.

Any idea whether this will be made possible?