Claude review: gpu: nova-core: convert Gsp::new() to use CoherentBox

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

Patch Review

Converts `libos` and `rmargs` initialization to use `CoherentBox` and `Coherent::init()` respectively.

**Ordering concern in `try_pin_init!`**: The new code initializes fields in this order:
```rust
loginit: LogBuffer::new(dev)?,
logintr: LogBuffer::new(dev)?,
logrm: LogBuffer::new(dev)?,
cmdq <- Cmdq::new(dev),
rmargs: Coherent::init(dev, GFP_KERNEL, GspArgumentsPadded::new(&cmdq))?,
libos: { ... libos.init_at(3, ..., rmargs)? ... libos.into() },
```
The `libos` initialization at element 3 references `rmargs` — this works because within `try_pin_init!`, earlier field bindings are accessible to later fields. The `rmargs` reference used in `LibosMemoryRegionInitArgument::new("RMARGS", rmargs)` is a `&Coherent<GspArgumentsPadded>`. This should work since `rmargs` has been initialized by that point in the macro expansion.

**`MessageQueueInitArguments` type alias change**: Changed from a newtype `struct MessageQueueInitArguments(bindings::MESSAGE_QUEUE_INIT_ARGUMENTS)` to a plain type alias `type MessageQueueInitArguments = bindings::MESSAGE_QUEUE_INIT_ARGUMENTS`. Then `impl MessageQueueInitArguments` adds a `new()` method returning `impl Init<Self>`. This is a clean simplification since the wrapper served no purpose beyond the constructor.

The `#[allow(non_snake_case)]` on the `new` function for `MessageQueueInitArguments` is needed since the `init!` macro generates bindings for the camelCase fields.

---
Generated by Claude Code Patch Reviewer