Claude review: gpu: nova-core: add CPU doorbell IRQ self-test

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

Patch Review

**Bug - test FAIL doesn't return error:**
```rust
if pass {
    dev_info!(pdev.as_ref(), "CPU doorbell self-test: PASS ...");
} else {
    dev_warn!(pdev.as_ref(), "CPU doorbell self-test: FAIL ...");
}

Ok(())
```
The function returns `Ok(())` regardless of whether the test passed or failed. This means a failing self-test only produces a `dev_warn` but probe succeeds. If this is intentional (don't block probe on selftest failure), it should be documented. If not, the failure path should return `Err(EIO)` to fail probe. The early-exit path for "doorbell bit already pending" *does* return `Err(EIO)`, making the inconsistency more apparent.

**Atomics ordering:** The handler uses `Relaxed` ordering for both `irq_count` and `doorbell_leaf_mask`:
```rust
self.irq_count.fetch_add(1, Relaxed);
self.completion.complete_all();
```
And the reader side:
```rust
let count = handler.irq_count.load(Relaxed);
let leaf_mask = handler.doorbell_leaf_mask.load(Relaxed);
```
Since `complete_all()` internally uses a completion which provides synchronization (the `wait_for_completion_timeout` acts as a barrier), the `Relaxed` ordering is actually fine here - the completion's internal synchronization guarantees visibility of the stores done before `complete_all()` to the thread that wakes from `wait_for_completion_timeout`. Good.

**Race window:** Between `handler.intr_ctrl.top().arm(bar)` and `handler.intr_ctrl.trigger(bar, DOORBELL_VECTOR)`, a spurious interrupt could arrive. The test only checks `count == 1`, so a spurious interrupt hitting the doorbell leaf would cause `count == 2` and the test would fail. This is unlikely but worth noting in a comment.

**Minor:** The `irq::Flags::TRIGGER_NONE` for the IRQ handler is correct since MSI is always edge-triggered and the trigger type is set by the MSI infrastructure.

---
Generated by Claude Code Patch Reviewer