fix(models): support llm-compressor NVFP4 MoE export variants

[Romeriz]

Fixes #410
Related: #390 (complementary -- FP8-attn/shared CT exports; see the #390 comment)

Summary

llm-compressor Qwen3.5-MoE NVFP4 exports come in layouts main could not convert
or serve: per-expert vs stacked per-layer routed experts, a GDN left bf16
or fp8 by the export's ignore list or a recipe: "all,gdn:fp8,-router",
format-only configs with no config_groups, native NVFP4 lm_head / embed
tokens, and single-file checkpoints without a shard index. Detect each variant
from the on-disk layout / recipe, build the matching offload banks (per-expert and
stacked loaders), and emit the state dict the model expects.

Changes

• models/config.py -- detect_compressed_tensors_nvfp4 falls back to the exact
  top-level format: nvfp4-pack-quantized when config_groups is absent;
  ModelConfig.gdn_quant (GDN out_proj quant, independent of attn_quant).
• models/nvfp4_banks.py -- _weight_map() (index OR single-file header);
  Nvfp4ExpertSourceSpec.stacked; load_nvfp4_stacked_expert_sources() reshapes each
  [E*rows, cols] bank tensor to [E, rows, cols] and broadcasts the per-projection
  layer-global (reciprocal) into the *_global banks. Same 6-bank dict as the
  per-expert loader, so the marlin/b12x repack and the offload cache are unchanged.
• models/qwen3_5_moe/config.py -- _ct_expert_groups_nvfp4 (+format fallback);
  _ct_recipe / _ct_ignored / _ct_gdn_nvfp4 read the export's ignore list and
  recipe string; _expert_quant recognizes CT-MoE -> nvfp4; CT lm_head_quant
  is nvfp4 unless the export skipped lm_head; gdn_quant wired.
• models/qwen3_5_moe/gdn.py, model.py -- GDN out_proj built from gdn_quant
  (defaults to attn_quant, preserving the modelopt/fp8 paths).
• models/qwen3_5_moe/weight.py -- _dequant_fp8_weight broadcasts per-output-row
  scales; per-expert CT + stacked CT expert specs and a spec selector that probes the
  weight map (index or header); _iter_weights_compressed_tensors now serves MoE
  checkpoints: skips per-expert and stacked experts (they ride the offload banks),
  dequantizes an fp8/bf16 GDN (fused in_proj), dequantizes NVFP4 embed_tokens,
  keeps lm_head native when the model wants it, and keeps NVFP4 attn / shared-expert
  native. Parallel bank read raises NotImplementedError for stacked experts (the
  serial read already saturates and the auto-pick already selects serial).
• tests/models/test_qwen3_5_moe_ct_nvfp4.py -- 16 weight-free tests: config
  detection (config_groups / format-only + recipe), recipe GDN override, per-row fp8
  dequant, per-expert + stacked spec and keying detection, per-expert and stacked
  single-file bank-loader placement with distinct gate_up/down globals, and full
  dense-pass fixtures for the AEON and Kwaipilot layouts.

Backwards compatibility

• Pure modelopt NVFP4 (nvidia/Qwen3.6-35B-A3B-NVFP4), block-fp8, and dense
  compressed-tensors NVFP4 (Qwen3.6-27B) take unchanged code paths.
• AEON-7 re-read after the change is identical to before (663-tensor state dict,
  bf16 GDN out_proj, native NVFP4 shared-expert/attn).
• Ornith-style CT exports (per-tensor-FP8 attn/shared-expert group + NVFP4 experts)
  are intentionally out of scope -- that is PR fix(models): better support for mixed-precision compressed-tensors NVFP4  #390; the two are complementary and
  overlap in shared plumbing (_CT_NVFP4_FUSE shared_expert fusion, _expert_quant,
  per-expert llm-compressor bank naming), so they should be reconciled (per-variant
  MoE-CT dense routing) rather than merged independently.

Tested on (real hardware)

• GPU: NVIDIA GeForce RTX 5060 Ti 16Gb, driver KMD Version: 610.57.04 CUDA UMD Version: 13.3 ; OS: DEbian 13 (trixie)
• FreeToken commit f8d6a56 on main (af71ba4); torch 2.11.0+cu130
• Checkpoints (all convert + serve):
  • AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4 (CT NVFP4, per-expert experts, GDN bf16,
    single 23 GB no-index file) -- FTW: 663 weight + 240 experts_bank, 20.96 GiB,
    quant_format: nvfp4
  • doth4580/Kwaipilot-KAT-Coder-V2.5-Dev-NVFP4-MIXED (format-only, stacked experts,
    gdn:fp8 recipe, NVFP4 lm_head/embed_tokens)
  • Ttimms/KAT-Coder-V2.5-Dev-REAP-50-NVFP4A16 (per-expert experts, GDN bf16, 128 experts)
• Commands: ft checkpoint --model <checkpoint_dir> --out <ftw_dir> then
  ft serve --model-path <ftw_dir>
• On main all three fail to convert or serve; on this branch they convert and serve,
  and the dense pass matches the model state dict exactly (AEON: 663 tensors,
  Kwaipilot: 665 tensors; 0 missing / 0 extra each).