Use pack and unpack in `join_nonshared_inputs`

[jessegrabowski]

Description

We now have symbolic helpers to do the operation here. It's basically the same as in this test

[ricardoV94]

It's a bit different because we use the initial_point dict as the source of shape information for unpack, which is always static

[jessegrabowski]

That should be fine though, we can call pack discarding the shape info, then pass the static shape info to unpack.

[AR21SM]

Hi @jessegrabowski, I’m interested in working on this issue. Please let me know if I can take it up.

[jessegrabowski]

I need to see if @ricardoV94 agrees that it's even something that should be done in principle

[ricardoV94]

Maybe? I did have some related changes to that function in https://github.com/pymc-devs/pymc/pull/7745/changes

And I think some tests started failing.

We can definitely use unpack to do the loop/split logic after we fix pymc-devs/pytensor#1835

I don't think we need to call pack just to get the type of the flat tensor. That part of the suggestion seems a tad wasteful.