Rigid assumption about Observations in acme.agents.tf.mcts

In `acme.agents.tf.mcts.types`, Observation is hinted to be a numpy ndarray
```python
# Assumption: observations are array-like.
Observation = np.ndarray
```
based on this, in `acme.agents.tf.mcts.acting`, the `MCTSActor._forward()` method hard-codes a `tf.expand_dims` call on the observation. This makes it impossible to pass nested structures as observations to MCTS.

To fix it, we can re-define this `_forward` method as:

```python
import tree
(...)
def _forward(self, observation):
# this is all they should have done in the first place
logits, value = self._network(tree.map_structure(lambda o: tf.expand_dims(o,axis=0), observation))
(...)
```
and solve the problem.

Is there anything I am missing?