…rward_qwen3_moe (#1749) (#1751)

Closes #1749. Pre-fix, `apr bench` against any MoE GGUF
(Qwen3-Coder-30B-A3B-Instruct etc.) routed through the dense
`forward_single_with_cache` path which calls `matmul_fused.rs:211`
on tensor names that don't exist on MoE models (the 3D `*_exps`
tensors are stored at different names than the 2D dense
`ffn_{gate,up,down}.weight` the dense path looks up). Result:
hundreds of parallel thread panics — `index out of bounds: len=0 but
index ≈ 91M`.

This PR adds MoE detection via `gguf.expert_count().is_some()` and
routes to the MoE-aware forward path:

  CPU:  realizar::gguf::OwnedQuantizedModel::forward_qwen3_moe
  CUDA: realizar::gguf::OwnedQuantizedModelCuda::forward_qwen3_moe_cuda

Both helpers do not currently expose a KV cache, so the bench runs
them **autoregressively with re-prefill** — each iteration runs full
forward over `prompt + previously-generated tokens` and appends the
argmax to the prompt for the next iter. O(N²) in N tokens but bounded
by `--max-tokens` (default 32).

This is intentionally a stop-gap to unblock M-GPU-MOE-3 PR-4
throughput measurement. True KV-cache MoE decoding is the actual
PR-4 work; this PR makes `apr bench` produce a real (if pessimistic)
tok/s number for MoE GGUFs instead of panicking.

## Empirical (lambda-vector RTX 4090, 2026-05-17)

  apr bench /home/noah/models/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf \
    --max-tokens 8 --warmup 1 --iterations 4 --json

  → total_time_ms: 87085 ; total_tokens: 4
  → 0.046 tok/s effective (auto-regressive re-prefill cost dominates)
  → 22.8s ttft, 22.2s p50, 32.5s p99

The 0.046 tok/s is the upper bound on what `apr bench` can currently
measure for MoE without KV cache. PR-4's job is to add the cache and
push this to ≥ 150 tok/s.

## What's in this PR

  crates/apr-cli/src/commands/bench_moe.rs (new):
    - `is_moe_gguf(&GGUFModel)` predicate
    - `run_gguf_moe_benchmark` — loads MappedGGUFModel + N
      Qwen3MoeQuantizedLayer descriptors + (optionally) wraps in
      OwnedQuantizedModelCuda, then dispatches to the CUDA or CPU
      bench helper.
    - `run_cuda_moe_benchmark` — autoregressive
      forward_qwen3_moe_cuda + greedy argmax decode.
    - `run_cpu_moe_benchmark` — autoregressive forward_qwen3_moe.

  crates/apr-cli/src/commands/bench.rs:
    + `include!("bench_moe.rs")` after the existing
      `include!("bench_safetensors.rs")` (same pattern as the other
      bench sub-files).

  crates/apr-cli/src/commands/benchmark.rs:
    + In `run_gguf_benchmark`, after parsing the GGUF and tokenising
      the prompt, check `is_moe_gguf(&gguf)`. If true, log the
      detection (`expert_count` + top-k) and tail-call
      `run_gguf_moe_benchmark`. Otherwise fall through to the
      existing dense path.

## What's NOT in this PR

  - True KV-cache MoE decoding (= M-GPU-MOE-3 PR-4 throughput target)
  - Streaming/per-token JSON output for MoE (existing JSON output
    works; just reflects the autoregressive re-prefill cost)
  - MoE bench for SafeTensors / APR formats (only GGUF MoE supported
    today; the other formats don't have MoE production paths in the
    realizar inference engine)

## Cross-refs

- #1583 M-GPU-MOE-3 — PR-4 throughput unblocks on this
- #1747 contract qwen3-moe-forward-gpu-v1 v1.7.2 (just merged) —
  L47 known-divergence + cascade pause point
- The MoE bench helpers reuse `forward_qwen3_moe[_cuda]` directly,
  which means PR-2 #1737's fp64 q6k_gemv acc is in effect; this
  bench measures *post-fp64-acc* throughput, not the pre-fix path.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>