[Question] question title

### ❓ Question

Hi guys.
In case of custom FeatureExtractor is it possible to teach some parts of it controlled but in terms of general Algorithm flow. I.e there's exact specific behavior is expected for the feature extractor to perform. So that it should not be trained as a black box along the `input - extractor - actor` flow but as a white box.

E.g. having environment that returns extra data within `info` result from a `step()` call (as it is not the observation it self)
```python
obs, reward, terminated, info, done = env.step(action)
```
So that is it possible to organize `feature_extractor` some how to use specific data from the `info` as a target at back propagation phase.

Or should such logic be trained just separately and used exclusively at inference mode within `feature_extractor `

Thank you!

### Checklist

- [x] I have checked that there is no similar [issue](https://github.com/DLR-RM/stable-baselines3/issues) in the repo
- [x] I have read the [documentation](https://stable-baselines3.readthedocs.io/en/master/)
- [x] If code there is, it is [minimal and working](https://github.com/DLR-RM/stable-baselines3/issues/982#issuecomment-1197044014)
- [x] If code there is, it is formatted using the [markdown code blocks](https://help.github.com/en/articles/creating-and-highlighting-code-blocks) for both code and stack traces.

Start with the custom FeatureExtractor documentation and the environment step() API described in the issue, then trace how observations and info are passed through training. No file, test, or entry point is named; the desired behavior and acceptance criteria need maintainer clarification before implementation.

Written by the indexing model from the issue text.