`maskless_step` and `maskless_backward_step` should throw an error if you don't return a tensor

Right now when writing environments, it's quite natural to manipulate the states object in place, and return the modified state. However the Env class expects a Tensor to be returned, not a States class instance.

E.g.,:

    def maskless_step(self, states: States, actions: Actions) -> TT["batch_shape", 1, torch.float]:
        states.tensor[..., 0] = states.tensor[..., 0] + actions.tensor.squeeze(-1)  # x position.
        states.tensor[..., 1] = states.tensor[..., 1] + 1  # Step counter.
        return states

throws a bug downstream,

[/usr/local/lib/python3.10/dist-packages/gfn/env.py](https://localhost:8080/#) in step(self, states, actions)
    158         #     new_not_done_states.masks = self.update_masks(not_done_states, not_done_actions)
    159 
--> 160         new_states.tensor[~new_sink_states_idx] = new_not_done_states_tensor
    161 
    162         return new_states

TypeError: can't assign a LineStates to a torch.FloatTensor

The fix is simply to return states.tensor.

GFNOrg / torchgfn

`maskless_step` and `maskless_backward_step` should throw an error if you don't return a tensor #145