feat(moe): support gsq q2_0 routed experts

[wenlive]

GSQ Q2_0 routed experts cannot currently bind to the MoE path. This adds an expert-only GGUF source for Qwen3.8-Flash-Next while preserving the HF trunk, shared-expert and PLE quantization schemes.

Related to #549. This does not implement complete GSQ GGUF model loading or IQ3_XXS, so it does not close that issue.

• Native 64-weight / 18-byte blocks, bounded complete-expert reads, and directory/geometry/receipt checks before bank allocation.
• Existing offload cache with a BF16/FP32 Triton vector kernel; initially TP=1, SiLU and pure GPU offload.
• Real miniature GGUF tests, independent Torch numerical references, slot remapping and CUDA graph replay; source contract and usage.

Draft for maintainer feedback on the external-source contract and reader placement. Maintainer design discussion, Linux testing, GPU testing with main's latest dependencies and full-model generation on this exact feature remain pending. No end-to-end performance claims.

Tested on Windows 11, Ryzen 9 9900X, 64 GB RAM, RTX 5090 D v2 24 GB, driver 610.88, Python 3.12.13:

$env:PYTHONPATH = "$PWD\python"
python -B -m pytest tests/moe/test_q2_0.py tests/moe/test_q2_source.py -q

• Torch 2.11.0+cu130 / triton-windows 3.6.0.post26: 16 passed.
• Torch 2.14.1+cpu: 11 passed, 5 CUDA cases skipped.
• Complete source read: all 48 x 512 experts / 33,973,862,400 packed bytes from ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF revision ed59f92082b1e93c0e96d60a8b11aab089b52f09; no inference or full bank allocation.

Commands and raw validation logs.

AI-assisted. These tests were executed by the assistant on my local hardware; a separate manual human rerun is not claimed.

[jason-fxz]

Thanks for the careful work here. #690 loads the ISTA-DASLab Qwen3.8-Flash-Next-GSQ-RCO IQ3_XXS GGUF directly, Q2_0 experts included, on llama.cpp's ggml-cuda kernels, and it runs end to end with fused and offload. It does not use the HF trunk plus GGUF experts approach of this PR.

#690 is the new GGUF infrastructure. Since it loads the whole GGUF file directly, we are not planning to merge a separate GGUF expert path, and this PR can be closed.