Claude review: gpu: nova-core: send UNLOADING_GUEST_DRIVER GSP command upon unloading

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

Patch Review

This is the core functional patch. Several observations:

**`shutdown_gsp` implementation**: The polling approach with `read_poll_timeout` for `LIBOS_INTERRUPT_PROCESSOR_SUSPENDED` (bit 31 of mailbox0) with a 5-second timeout and 10ms poll interval is reasonable.

**`PowerStateLevel::is_power_transition`**: Clean design — only `Level0` is a full unload, others preserve state:
```rust
pub(crate) fn is_power_transition(self) -> bool {
    self != PowerStateLevel::Level0
}
```

**`UnloadingGuestDriverReply`**: The reply type uses `type Message = ()`, meaning the GSP sends no payload in its reply. This is consistent with OpenRM's behavior for this RPC.

**`unbind` in `driver.rs`**: The `unbind` callback receives `dev: &'bound pci::Device<kernel::device::Core>` — note it's `device::Core`, not `device::Bound`. This is intentional since at unbind time the device is being unbound. The `gpu.rs` `unbind` method correctly calls `pdev.as_ref()` to get a `device::Device<device::Bound>` for the `unload` call — wait, actually `pci::Device<Core>` would yield `device::Device<Core>` via `as_ref()`. Let me check:

```rust
pub(crate) fn unbind(&self, pdev: &'bound pci::Device<device::Core>) {
    let _ = self
        .gsp
        .unload(pdev.as_ref(), self.bar, &self.gsp_falcon)
```

And `unload` takes `dev: &device::Device<device::Bound>`. This would be a type mismatch if `pdev.as_ref()` returns `&device::Device<Core>` rather than `&device::Device<Bound>`. However, the Device HRT framework may provide the right conversions here. This could be a compile-time issue but since the author presumably tested this, the types likely work out.

**Minor**: The `#[expect(unused)]` on the `PowerStateLevel` enum is good practice since `Level3` and `Level7` aren't used yet.

---

---
Generated by Claude Code Patch Reviewer