Handle multivariate responses with HSGP

[tomicapretto]

This PR does two things.

• Main: Make it possible to use HSGP with multivariate responses, such as a multinomial likelihood.
• Adjacent: Make multivariate families use two dims. So far, it created an extra dimension which was not equal to the response dimension.

There's something I would like to clarify about HSGP with multivariate responses. With this PR it's possible to do something like

formula = "c(y1, y2, y3) ~ 0 + hsgp(x, m=30, c=2)"
hsgp_model = bmb.Model(formula, df, family="multinomial")

# Setting aliases for a nicer graph
hsgp_model.set_alias({"c(y1, y2, y3)": "result"})
hsgp_model.set_alias({"hsgp(x, m=30, c=2)": "hsgp"})
hsgp_model.build()
hsgp_model.graph()

and the graph will look like:

See all the dimensions of the response share the same priors for hsgp_sigma and hsgp_ell. Theoretically, it's possible to use a different coefficient for each response dimension. However, Bambi does not support that, and after some thought I decided it is fine that way.

The implementation is very complicated already, and it would be much more complicated if we decided to handle this. There's the by (and share_cov) argument in hsgp(), which could make us think we can use it for this purpose. However, that argument expects categories to be values within a given variable. In the multivariate family case, the dimensions are different columns (i.e. y1, y2, and y3 in the example above). So, at least, we would need a special way to tell Bambi to handle things differently in this special case.

On top of that, a multivariate model with an HSGP is already a fairly complex model. To have a more granular control of HSGP, one should use a PPL like PyMC.

I'm open to change my mind in the future, but for now, I think this is good enough.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 69.23077% with 4 lines in your changes missing coverage. Please review.

Project coverage is 89.33%. Comparing base (516d7bd) to head (59f2f7a).
Report is 7 commits behind head on main.

Files with missing lines	Patch %	Lines
bambi/backend/terms.py	60.00%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #856      +/-   ##
==========================================
- Coverage   89.71%   89.33%   -0.39%     
==========================================
  Files          47       47              
  Lines        3997     4030      +33     
==========================================
+ Hits         3586     3600      +14     
- Misses        411      430      +19     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[AlexAndorra]

LGTM @tomicapretto , thanks a lot for adding this!
I'd vote to allow share_cov=False also for these cases, but I understand it's hard to implement. I'm happy to contribute this though, if you think it's not too big