fix(mir): make BackLink parent removal precise

[phrwlk]

Describe your changes

BackLink implements PartialEq as always true so that backlink fields are ignored in derived equality and hashing. However,
remove_parent in MIR ops compared BackLink values directly:

self.parents.retain(|p| *p != parent.clone().into());*

Because of the always-true PartialEq, this predicate was always false and effectively cleared the entire parents vector instead of removing only the requested parent.

This change updates all remove_parent implementations for ops that use parents: Vec<BackLink> to compare against the upgraded Link via to_link(), keeping BackLink’s structural Eq/Hash semantics intact while making parent removal semantically correct.

[AlgofootPrint]

Review: Fix is incomplete. Link::PartialEq compares by value, not pointer identity @phrwlk

I tested this PR locally and the included test add_remove_parent_removes_only_target fails:

thread 'ir::nodes::ops::add::tests::add_remove_parent_removes_only_target' panicked
assertion left == right failed
left: 0
right: 1

The root cause diagnosis is correct — BackLink::PartialEq always returning true breaks retain. However, the proposed
fix still has a problem.

Switching to p.to_link() and then link != parent uses Link::PartialEq, which compares by value (self.link ==
other.link → Rc<RefCell> → inner value equality). So if two different owners happen to hold the same value (as in
the test: both are Owner::None(span)), they compare as equal and both get removed, leaving parents.len() == 0.

The fix needs pointer identity, using the already-available get_ptr() method on Link:

fn remove_parent(&mut self, parent: LinkSelf::Parent) {
self.parents.retain(|p| match p.to_link() {
Some(link) => link.get_ptr() != parent.get_ptr(),
None => true,
});
}

This uses Rc::as_ptr under the hood, which correctly distinguishes different allocations regardless of value
equality. The same fix needs to be applied to all 16 op files.

The test itself is well-structured and proves the regression correctly but it just needs the implementation to match.