Claude review: gpu: nova-core: gsp: support large RPCs via continuation record

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

Patch Review

This is the core patch. The design is solid:

- `ContinuationRecord<'a>` wraps a `&[u8]` slice, uses `MsgFunction::ContinuationRecord`, and has `type Command = ()` (zero-sized header). This is correct since continuation records are pure payload.

- `ContinuationRecords` provides a lending-iterator pattern via a custom `next()` method rather than implementing `Iterator`, which is the right approach since `ContinuationRecord` borrows from the parent.

- `SplitState<C>` cleanly separates the single vs. split cases. The associated constant `MAX_FIRST_PAYLOAD` properly accounts for the command header:

```rust
const MAX_FIRST_PAYLOAD: usize = MAX_CMD_SIZE - size_of::<C::Command>();
```

- `SplitCommand<C>` delegates `init()` to the original command but replaces the variable payload with the pre-filled staging buffer.

**Observation on the send_command flow:** For split commands, `SplitState::new()` calls `command.init_variable_payload()` to fill the staging buffer. Then `send_single_command` calls `SplitCommand.init_variable_payload()` which copies from the staging buffer into the command queue. This means the payload is initialized once and copied once (for the first chunk) — which is correct and matches the cover letter's description.

**Observation on ContinuationRecord being sent through send_single_command:** The `allocate_command` size check uses `size_of::<GspMsgElement>() + size`. For ContinuationRecord, the size is `0 (Command=()) + data.len()`, and `data.len() <= MAX_CMD_SIZE`, so `size_of::<GspMsgElement>() + data.len() <= GSP_MSG_QUEUE_ELEMENT_SIZE_MAX`. The check passes correctly.

**Minor style observation:** The trailing `//` in imports is a rustfmt formatting workaround:

```rust
ContinuationRecord,
SplitState, //
```

This is a known pattern but could be confusing to newcomers. Not a blocking issue.

---
Generated by Claude Code Patch Reviewer