Add zero-sum encoding option for categorical predictors

[Chirag3841]

Thank your for opening a PR!

Before you proceed, please check the following notes.

• Make sure you describe the changes introduced in the PR.
• If this PR addresses an existing issue, please mention it here.
• Check our Guidelines for Contributing.
  In particular, pay attention to:
  • Contributing steps.
  • Pull request checklist.

Issue #941
Description

This PR adds support for an alternative categorical encoding strategy using a zero-sum constraint via pm.ZeroSumNormal.
Users can now choose between "reference" (default) and "zero-sum" encoding through the new categorical_encoding argument in bmb.Model.

Changes done
Added categorical_encoding argument to bmb.Model
Implemented "zero-sum" encoding for categorical predictors in PyMC backend using pm.ZeroSumNormal
Added unit test to verify model builds correctly with categorical_encoding="zero-sum"

[tomicapretto]

@Chirag3841, thanks for the initiative to start this contribution. I'm afraid supporting zero-sum encoding in Bambi is a way more complex endeavor than what this PR attempts to do. For instance, we'd need to modify the model-building library, formulae, to return non-full rank matrices in cases we want to use ZeroSum normal. And then we'd need to modify multiple parts of Bambi to account for that change. The addition of a model-level argument is not a suitable solution, as the encoding of each categorical covariate is controlled at the covariate level.