Support for custom environments

Hello, I'd like to ask you a few questions：
I am currently applying MARL to multi-drone cooperative navigation tasks, and my environment is a custom setup based on [https://github.com/utiasDSL/gym-pybullet-drones](url).The environmental observations include state vectors and depth maps, as shown below:
```
return spaces.Dict({f"agent_{i}": spaces.Dict({"state": spaces.Box(low=obs_lower_bound,
high=obs_upper_bound,
dtype=np.float32
),
# depth image
"dep": spaces.Box(low=0.0,
high=1.0,
shape=(self.IMG_RES[1],
self.IMG_RES[0],
1),
dtype=np.float32
),
}) for i in range(self.NUM_DRONES)})
```
I'd like to know if BenchMARL is compatible with my environment? If so, I might need to use MLP and CNN to process the state vector and depth map respectively. What modifications would I need to make to BenchMARL?

No files or tests are named. Start by reviewing BenchMARL's environment integration and observation/model handling, then check whether the custom Dict observation with state vectors and depth maps can use separate MLP and CNN processing. Done means establishing compatibility and documenting or scoping the required modifications.

Written by the indexing model from the issue text.