feat: add scoped physical expression traversal

[rluvaton]

Mostly tests

Which issue does this PR close?

Relates to #21231

Rationale for this change

CaseWhen (and other expression utilities) discover input columns by traversing the expression tree via children(). This breaks when a child expression introduces a new naming scope — for example, a lambda expression like array_transform(list, item -> item + external_col) where item is bound to a different schema than the outer expression.

Today, the only workaround is to not return scoped children from children(), but that breaks other legitimate use cases (e.g. finding all expressions in a tree, wrapping every expression with debug/timing, etc.). We need a way to distinguish "same-scope children" from "all children".

What changes are included in this PR?

New ScopedTreeNode trait (datafusion/common/src/tree_node.rs):

• Extends TreeNode with the concept of scopes — a subtree boundary that traversals should not cross unless intended
• Provides scoped variants of all standard traversal methods: apply_in_scope, visit_in_scope, rewrite_in_scope, transform_down_in_scope, transform_up_in_scope, transform_down_up_in_scope, exists_in_scope
• Includes DynScopedTreeNode for Arc<dyn T> blanket implementations

New methods on PhysicalExpr trait:

• children_in_scope() — returns only children that share the same input schema as the current expression. Defaults to children(), so no breaking change for existing implementations
• with_new_children_in_scope() — reconstructs the expression with new scoped children. Defaults to with_new_children()

Updated expression utilities to use scoped traversal:

• CaseWhen column projection now uses apply_in_scope / transform_down_in_scope so it no longer traverses into lambda bodies or other new-scope children
• collect_columns and related helpers in physical-expr/src/utils use scoped traversal where appropriate

No changes to non-scoped traversals — TreeNode::visit, display/formatting, and other tree walks that intentionally need the full tree continue to use children().

Are these changes tested?

Yes:

• Unit tests for ScopedTreeNode traversal on TestTreeNode with mixed scoped/non-scoped children
• Tests verifying CaseWhen does not incorrectly project columns that belong to a different scope (e.g. lambda variables)
• Existing tests continue to pass

Are there any user-facing changes?

Yes, but no breaking changes:

• New public trait ScopedTreeNode in datafusion-common
• New default methods children_in_scope() and with_new_children_in_scope() on PhysicalExpr
• Existing implementations that don't override these methods get the current behavior automatically

[rluvaton]

@gstvg how would that affect your lambda support, would that make it easier or harder to implement column capture?

note that in column capture you preferably should only keep the columns that are used to avoid a lot of copies

[alamb]

I think this API is confusing because "scope" isn't something that is defined

I think a beter description of what this API is doing is "children in the outermost scope"

All built in expressions have the (implicit) property that their children are evaluated in the same input schema. I don't see how CASE is any different in theory (though I understand it is what is causing you current problems)

[alamb]

I also think if we add thsi API it is easy to misue and you'll be chasing down corner case bugs with other expression types that have the same problem

[gstvg]

@gstvg how would that affect your lambda support, would that make it easier or harder to implement column capture?

I believe it wouldn't make much difference, that's okay

note that in column capture you preferably should only keep the columns that are used to avoid a lot of copies

Thanks, I will make sure to implement that