Custom likelihood variables

[jessegrabowski]

This was originally @ricardoV94's PR, but I clobbered it when I moved the repo. Sorry.

[ricardoV94]

there were some PR comments in the old PR but I guess they are lost forever

[jessegrabowski]

there were some PR comments in the old PR but I guess they are lost forever

I have them in pycharm cache. I addressed them in my follows ups: removed build (made a thin helper that goes in the user-facing functions), leaned into GPData (though you had already did that, you mooted your own comment).

You also flagged the need for more testing (you only tested clone), that's still a to-do.

[ricardoV94]

it was more the motivation for my changes on top of yours, and whether it makes sense for custom type or we'd be okay with a functional api like pm.logp(rv, value) instead of rv.logp(value). For Likelihood we can be cute and add these methods, but I dunno what other "GP objects" are there around. I suppose there are some latent GPs? with similar methods, and I dunno if there have actual tensor meaning right now (the f or whatever)?

[jessegrabowski]

Personally i think it's great, it solves the problem very cleanly in an idiomatic pytensor way