Claude review: rust: auxiliary: generalize Registration over ForLt

[Claude Code Review Bot <claude-review@example.com>]

Patch Review

`Registration<T>` becomes `Registration<'a, F: ForLt>`. The key split:

```rust
+    pub unsafe fn new_with_lt<E>(
+        parent: &'a device::Device<device::Bound>,
+        ...
+        data: impl PinInit<F::Of<'a>, E>,
+    ) -> Result<Self>
```

vs the safe version:

```rust
+    pub fn new<E>(...) -> Result<Self>
+    where
+        F::Of<'a>: 'static,
+    {
+        unsafe { Self::new_with_lt(parent, name, id, modname, data) }
+    }
```

The `new_with_lt` is unsafe because `mem::forget()` could prevent `Drop` from running, leaving borrowed references dangling. The `new()` is safe when data is `'static` because there are no borrowed references to dangle.

The `registration_data` method uses `ForLt::cast_ref` to shorten the stored `'static` lifetime:

```rust
+        let pinned: Pin<&F::Of<'static>> = unsafe { wrapper.map_unchecked(|w| &w.data) };
+        Ok(unsafe { Pin::new_unchecked(F::cast_ref(pinned.get_ref())) })
```

The `PhantomData<(fn(&'a ()) -> &'a (), F)>` in Registration ensures `'a` is invariant (via the function pointer type), which is correct -- the registration must not outlive the scope it was created in.

The `Registration` no longer wraps in `Devres` -- it's now directly held by the driver. The `Send`/`Sync` impls use `for<'a> F::Of<'a>: Send` instead of `T: Send + Sync`. Dropping `Sync` requirement from the `Send` bound seems intentional -- a `Registration` is moved, not shared.

The transmute from `RegistrationData<F::Of<'a>>` to `RegistrationData<F::Of<'static>>` is sound because lifetimes don't affect layout.

---
Generated by Claude Code Patch Reviewer