Claude review: drm: nova: convert to use DRM registration data

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

Patch Review

This is the demonstration patch. Nova moves the auxiliary device reference from `drm::Device` data into `RegistrationData`:

```rust
-pub(crate) struct NovaData {
-    pub(crate) adev: ARef<auxiliary::Device>,
+pub(crate) struct NovaData<'bound> {
+    pub(crate) adev: &'bound auxiliary::Device<Bound>,
 }
```

And the driver trait:
```rust
-    type Data = NovaData;
-    type RegistrationData = ForLt!(());
+    type Data = ();
+    type RegistrationData = CovariantForLt!(NovaData<'_>);
```

**Key change**: `Data` becomes `()` (no per-device data accessible without UnbindGuard) and `RegistrationData` becomes `CovariantForLt!(NovaData<'_>)` containing the borrowed `&'bound auxiliary::Device<Bound>`.

**`new_with_lt()` usage**:
```rust
+        let drm = drm::UnregisteredDevice::<Self>::new(adev, Ok(()))?;
+        let reg_data = NovaData { adev };
+        // SAFETY: We never bypass the destructor of `reg`.
+        let reg = unsafe { drm::Registration::new_with_lt(adev.as_ref(), drm, reg_data, 0)? };
```

The `new_with_lt()` is unsafe because the caller must not `mem::forget()` the Registration (which would cause use-after-free of the borrowed `adev` reference after the bus driver unbinds). The safety comment "We never bypass the destructor of `reg`" is correct — `reg` is stored in `Nova<'bound>` which is the bus driver's `Data<'bound>`, and the bus framework guarantees drop is called.

Note that `drm::UnregisteredDevice::new(adev, Ok(()))` passes `Ok(())` as the data init — this works because `PinInit<(), Error>` is implemented for `Result<()>`. Slightly unusual but correct.

**In `get_param()`**:
```rust
-        let adev = &dev.adev;
-        let parent = adev.parent();
-        let pdev: &pci::Device = parent.try_into()?;
+        let parent = reg_data.adev.parent();
+        let pdev: &pci::Device<Bound> = parent.try_into()?;
```

Now the PCI device reference is `&pci::Device<Bound>` instead of `&pci::Device` (unbound). This is a type-level improvement — the bound context guarantees the PCI device is still bound, which is correct within the `drm_dev_enter()` critical section.

Overall: this is a well-structured, carefully thought-out series that addresses a genuine soundness concern. The SRCU-based safety argument is correct (verified against the C `drm_dev_unplug()` implementation), and the HRTB closure pattern for lifetime-erased data access is the standard Rust approach for this kind of problem. The incremental patch structure makes it easy to review each step.

---
Generated by Claude Code Patch Reviewer