ENH: lift_subqueries scheduler pass

[mtsokol]

This PR:

• introduces lift_subqueries default scheduler pass
• adds pre-commit setup
• lints existing files according to ruff rules

[mtsokol]

In Finch.jl arguments(node) can be also called on non-tree IR instances. I think this should be implemented the same way here: Return empty list by default and is_expr=True IRs override it returning actual children.

[willow-ahrens]

I have mixed thoughts on this. Implementing children(node) for literal feels dangerous because x.make_term(x.head(), x.children()) is not equal to x. What do you think Mateusz? @wraith1995 do you have thoughts on this one? Maybe as long as we don't also implement make_term for this IR node type it's okay.

[mtsokol]

I'm also hesitant about this, so I followed Finch.jl:

julia> using Finch

julia> sq = Finch.subquery("lhs", "arg")
Finch Logic: (lhs = arg)

julia> imm = Finch.immediate("imm")
Finch Logic: imm

julia> Finch.arguments(sq)
2-element Vector{Finch.FinchLogic.LogicNode}:
 Finch Logic: lhs
 Finch Logic: arg

julia> Finch.arguments(imm)
Finch.FinchLogic.LogicNode[]

[mtsokol]

Now I would prefer to have it as:

    def children(self):
        """Returns the children of the node."""
        raise Exception(f"`children` isn't supported for {self.__class__}")

as a base implementation.

[wraith1995]

I have two thoughts:

1. Perhaps more thought should be given to how the language interacts with things outside of it - in an ideal world, nothing outside the IR except for constants would be in the IR or some process would remove such references and replace them with something else. Perhaps, this would fix the problem of make_term? If you have tensors in the expressions, perhaps they should be pre-processed out before anything else happens.
2. I generally support that IR instances should have a uniform set of methods so I don't think either should not implement anything. But perhaps returning [] is a bad idea - a pythonic way to do this would be to return None and handle the consequences downstream. Alternatively, you could have make_term fail on expressions by raising an exception. Either way, it requires code in various places to catch this.

[wraith1995]

3. I should add I believe (2) in part because I would want to move these methods to other places via just one big case statement.

[willow-ahrens]

I agree with 1, and I think that from the point of the view of the IR, everything that isn't is_expr is basically different kinds of constants. Mateusz, let's go forward with your Error-related approach for now. I think Teo's return-none approach is similar to if x.is_expr() x.children else None, which would need an if-branch anyway. We'll use Mateusz' approach for now since we're still translating from Finch, but I don't think that changing it to Teo's approach later would be a big issue.

[mtsokol]

👍 I made it an Exception then!

[mtsokol]

We should use tuple explicitly. We expect these LogicNodes to be hashable, and non-hashable structures like list are also Iterable.

[willow-ahrens]

Good call!

[mtsokol]

Renamed it to follow Finch.jl subquery node.

[willow-ahrens]

LGTM!