Fix Qwen3.6 (qwen3_5 / qwen3_5_moe) on Metal: RMSNorm, AFQ, lm_head, hybrid KV cache

[sergey-scherbina]

Adds support for Qwen3.6 - both the dense qwen3_5 and the MoE qwen3_5_moe
(35B-A3B) - and makes them numerically correct on Metal.

Root causes fixed

• RMSNorm +1 convention. GemmaRmsNorm::new bakes weight = on_disk + 1.0, but
  the sanitized mlx-community/Qwen3.6-* checkpoints ship raw RMSNorm weights
  (MLX's should_shift_norm_weights is false for these: conv1d (…,4,1), no MTP). The
  +1 over-scaled every norm (~2.1x) and the SiLU MoE compounded it into a ~14x
  experts blow-up and an over-peaked router. Fixed with an unshifted norm at the
  affected sites.
• AFQ packed loading / lm_head for the MLX-quantized checkpoints.
• Hybrid KV cache: full-attention layers use the paged KV pool; the
  linear-attention (GatedDeltaNet) layers carry a recurrent state. The cache/config
  plumbing wires both per layer_types.

The MoE experts ship pre-fused as switch_mlp.{gate,up,down}_proj; the existing
FusedExperts path loads them via afq_packed_linear_b and applies router weights
correctly - no layout change was needed (verified, was a red herring).

Correctness

"Hello" renders identically to mlx_lm (embedding byte-for-byte, all layer
last-position norms within bf16 rounding, identical top-1 logit), and greedy
generation begins identically. Verified on Qwen3.6-27B-4bit and Qwen3.6-35B-A3B-4bit
on Apple Silicon Metal.

Notes

• Rebased onto current main (clean).
• This PR is now scoped to Qwen3.6 only. It was split out of a larger branch into
  reviewable pieces; the zero-element KV buffer fix is a prerequisite for running
  this on Metal. Suggested merge order: zero-buffer + engine-reap → this → chunked-prefill.

Scope

gdn/*, vision_models/qwen3_5{,_moe}/*, paged_attention/config.rs, layers.rs,
mistralrs-quant/{afq,lib} - 12 files, +614/-74.

---

Part of splitting the Qwen3.6 work into focused, reviewable PRs:

• fix(metal): back zero-element KV cache buffers with a shared placeholder #2206 - zero-element KV cache buffer (Metal prerequisite for Fix Qwen3.6 (qwen3_5 / qwen3_5_moe) on Metal: RMSNorm, AFQ, lm_head, hybrid KV cache #2201)
• fix(engine): drop disconnected sequences before the prefill pass #2207 - reap disconnected sequences before prefill (independent)
• Fix Qwen3.6 (qwen3_5 / qwen3_5_moe) on Metal: RMSNorm, AFQ, lm_head, hybrid KV cache #2201 - Qwen3.6 (qwen3_5 / qwen3_5_moe) model support
• feat(metal): paged chunked prefill, env-tunable with a safe chunk-size floor #2208 - Metal paged chunked prefill

Suggested merge order: #2206 + #2207 -> #2201 -> #2208.

[github-actions]

Code Metrics Report

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 C Header                 23         4454         3116          790          548
 CSS                       3          281          252            5           24
 CUDA                    119        23575        19136         1696         2743
 Dockerfile                1           38           21            8            9
 HTML                      2           27           27            0            0
 JavaScript                3          392          387            2            3
 Jinja2                    7          694          656            5           33
 JSON                     26         9360         9357            0            3
 Makefile                  1            6            5            0            1
 MDX                       1          149            0          133           16
 Metal Shading Lan|       37        14287        11284         1136         1867
 PowerShell                1          357          276           33           48
 Python                  131        10342         8515          460         1367
 Shell                     2          549          379          101           69
 Plain Text                3         3723            0         2413         1310
 TOML                     29         1388         1211           41          136
 TypeScript               11         1607         1371           66          170
 YAML                      3           25           23            2            0
─────────────────────────────────────────────────────────────────────────────────
 Jupyter Notebooks         3          122           83           23           16
 |- Markdown               1           60           30           22            8
 |- Python                 1          122          113            1            8
 (Total)                              304          226           46           32
─────────────────────────────────────────────────────────────────────────────────
 Markdown                129         9703            0         6648         3055
 |- BASH                  61          600          520           47           33
 |- Dockerfile             2            5            5            0            0
 |- JSON                  18          700          700            0            0
 |- PowerShell             3            5            5            0            0
 |- Python                25          830          722            5          103
 |- Rust                  15          437          382            1           54
 |- TOML                  10          124           98            3           23
 |- YAML                   1           13           13            0            0
 (Total)                            12417         2445         6704         3268
─────────────────────────────────────────────────────────────────────────────────
 Rust                    625       270388       239956         5864        24568
 |- Markdown             397         9504          452         7882         1170
 (Total)                           279892       240408        13746        25738
─────────────────────────────────────────────────────────────────────────────────
 Svelte                   18         1831         1696           50           85
 |- CSS                    1            4            4            0            0
 |- JavaScript            18          876          727           24          125
 (Total)                             2711         2427           74          210
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                  1178       366578       301522        27461        37595
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━