refactor(forward): lift attention_layer_with_cache helper (M32d Day 1 prep, #1830 PR-1 of 4)

[noahgift]

Summary

Pure refactor of the dense KV-cache path's per-layer attention sub-block (steps 2a–2g) into a private helper method on `OwnedQuantizedModel`. Behavior is bit-identical to the previous inline version.

First of 4 PRs implementing #1830 (M32d KV cache for qwen3_moe inference path) per `docs/specifications/m32d-moe-kv-cache-scope.md`.

Why

M32d needs to add `forward_single_qwen3_moe_with_cache` that mirrors `forward_single_with_cache`'s structure but swaps the dense FFN block for MoE expert dispatch. The attention block is reusable as-is — extracting it into a shared helper eliminates the future copy-paste.

PR layout

#	What	Risk
1 (this)	Extract `attention_layer_with_cache` helper from dense path	Medium — must not regress dense KV cache
2	Extract `moe_ffn_layer` helper from `forward_qwen3_moe.rs`	Medium — per-expert routing logic
3	Add `forward_single_qwen3_moe_with_cache` + wire `run_qwen3_moe_generate`	Low — composition of helpers
4	`moe_kv_cache_equivalence` integration test (env-gated, `#[ignore]`)	High — float-equivalence is hard

PR 1 is the lowest-risk prep step. The 34 existing `single_tests` must remain green.

What changed

• `crates/aprender-serve/src/gguf/inference/forward/debug.rs` lines 477-589 (the per-layer attention sub-block inside `forward_single_with_cache`'s for-loop) lifted verbatim into a new private method `attention_layer_with_cache`.
• Inline block replaced with a single call to the helper.

The helper signature follows the spec doc exactly:

```rust
fn attention_layer_with_cache(
&self,
hidden: &mut Vec,
layer: &OwnedQuantizedLayer,
layer_idx: usize,
cache: &mut OwnedQuantizedKVCache,
position: usize,
attn_out_buffer: &mut [f32],
use_rmsnorm: bool,
) -> Result<()>
```

ALL existing behaviors preserved: PMAT-260 debug traces, GH-278 LayerNorm bias branch, GH-479 Qwen3 per-head QK RMSNorm, RoPE skip for absolute-position models, GQA expansion for empty-cache first-token, CORRECTNESS-013 attention output trace, fused_matmul output projection, attn_output_bias add, residual.

Test plan

• `cargo check -p aprender-serve --lib --features cuda` → clean
• `cargo test -p aprender-serve --lib --features cuda gguf::inference::forward::single` → 34 passed / 0 failed / 1 ignored including:
  • `test_forward_single_with_cache_q8k_path`
  • `test_forward_single_with_cache_gqa_256`
  • `test_forward_single_sequential_decode_10_tokens`
  • `test_forward_single_with_scratch_multi_token`
  • `test_forward_single_with_scratch_q8k_multi_position`

The multi-position + sequential decode tests are the load-bearing regression gate — they exercise the same attention path repeatedly with growing cache state.

Cross-refs

• Issue: M32d: KV cache for qwen3_moe inference path (engineer-driven, 1-2 week) #1830 (M32d KV cache for qwen3_moe inference path)
• Spec: `docs/specifications/m32d-moe-kv-cache-scope.md`
• Contract gate (downstream, blocked by full M32d): FALSIFY-QWEN3_MOE_SERVE_DISPATCH_V1_004 in `contracts/qwen3-moe-serve-dispatch-v1.yaml`
• Expected speedup once full M32d ships: ~30× (0.5 tok/s → 5-15 tok/s on Qwen3-Coder-30B-A3B)

🤖 Generated with Claude Code

[noahgift]

Superseded by #1832 — operator flipped from Option (b) engineer-driven to Option (a) in-session and shipped the full M32d KV cache feature monolithically (without the helper extraction this PR proposed). #1832 ships forward_single_qwen3_moe_with_cache + cache-aware run_qwen3_moe_generate + moe_kv_cache_equivalence test + m32d_perf bench with empirical 19× speedup. Closing this prep refactor PR; not needed once #1832 lands.

This is the standard chain-PR squash-leapfrog cleanup pattern (memory: feedback_chain_pr_squash_leapfrog.md).