Claude review: gpu: nova-core: use I/O projection for cleaner encapsulation

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

Patch Review

This is the practical payoff: nova-core's `gsp_mem` module (which was a workaround for accessing shared DMA memory fields) is deleted entirely, and replaced by `io_project!` + `io_read!`/`io_write!` calls. The `GspMem` and `Msgq` structs can become private again, removing the `pub(super)` visibility that was only needed for the workaround module.

The `PteArray::init` method now takes a `View` directly:
```rust
fn init<T: KnownSize + ?Sized>(
    view: io::View<'_, Coherent<T>, Self>,
    start: DmaAddress,
) -> Result<()> {
```

This is much cleaner than the old slice-based approach. The fence operations are preserved correctly.

One notable change: initialization of `cpuq.tx` and `cpuq.rx` now uses `CoherentBox` (exclusive access) before converting to `Coherent`:
```rust
let mut gsp_mem = CoherentBox::<GspMem>::zeroed(dev, GFP_KERNEL)?;
gsp_mem.cpuq.tx = MsgqTxHeader::new(MSGQ_SIZE, RX_HDR_OFF, MSGQ_NUM_PAGES);
gsp_mem.cpuq.rx = MsgqRxHeader::new();
let gsp_mem: Coherent<_> = gsp_mem.into();
```

This is better than the old `dma_write!` approach since the headers are fully initialized before the DMA region becomes shared.

---
Generated by Claude Code Patch Reviewer