feat(models): add nvidia/Qwen3.8-Flash-Next-FP4-Local checkpoint support

[chrisqianz]

Adds support for serving nvidia/Qwen3.8-Flash-Next-FP4-Local on top of the existing qwen4_exp (Qwen3.8-Flash-Next NVFP4) framework. The two checkpoints share the architecture; FP4-Local differs only in byte format and quant coverage:

• Routed experts: per-16 NVFP4 scales stored in the CSF slab encoding (weight_scale.nvfp4_csf_fixed U8 slabs + weight_scale.nvfp4_csf_exceptions U32 records). decode_csf_scales expands them on CPU to plain per-16 E4M3 so the existing expert reader and triton kernel pipeline are unchanged. The expert spec is selected automatically from the weight_map (presence of .nvfp4_csf_fixed).
• PLE table (layers.1.ple): 128 shards of packed e2m1 U8 rows + per-shard E4M3 scale rows + one global F32 scale. Loaded as packed e2m1 (no BF16 materialization, 25.5 GB pinned instead of ~102 GB) and dequantized by the triton PLE gather kernel (IS_FP4 path). --ple-backend disk falls back to the pinned backend automatically for packed tables (the disk row store streams native rows only).
• Dense projections: FP4-Local quantizes the text projections (GDN in_proj_qkv/z, self_attn q/k/v/o, index_qk) to MXFP8 per-32 and the vision MLP to W4A16 NVFP4, while the non-Local release leaves them bf16. The reader maps checkpoint scale names to the engine's buffer names per quant scheme (weight_scale -> weight_scale_inv, weight_scale_2 -> weight_global) and expands the scalar W4A16 global to the model's [out] fp16 buffer.
• Checkpoint layout: FP4-Local keeps its 21 shards in a tensors/ subdirectory referenced by the index; iter_weight_files falls back to the index's weight_map when the top-level glob finds nothing (also fixes any future index-only HF layout).
• The parallel expert loader's dtype table gains U16/U32/U64 (the CSF exceptions stream is U32).

Committer ran the model end to end with ft serve and it is serving real requests (radix cache hits, ~5k tok/s prefill, ~80 tok/s decode).

Hardware

• NVIDIA RTX 5090 D, 32 GB
• Driver / CUDA: torch 2.11.0+cu130, transformers 5.16.1
• FreeToken: this branch rebased onto main (9b68e00)

Checkpoint

• nvidia/Qwen3.8-Flash-Next-FP4-Local (21 shards in tensors/, ~96 GB)

Exact command

ft serve --model <path> --ple-backend pinned   --moe-strategy offload --moe-cache-auto   --max-running-requests 1 --kv-reserve-tokens 131072   --moe-prefill-hit-d2d --host 0.0.0.0   --served-model-name Qwen3.8-Flash-Next-FP4-Local --enable-cache-report

Tests

uv run pytest tests/models/qwen4_exp/ tests/models/test_models_loader.py

New coverage: CSF decode round-trip, expert spec auto-selection, PLE packed-table load, e2m1 kernel vs LUT, dense scale renames, W4A16 global expansion, indexed-subdirectory weight discovery, PLE backend fallback.