fix(cuda): support GLM IQ2_XXS routed down projection

[hy-sde]

Summary

Fix GLM-5.2 all-IQ2_XXS routed-MoE execution on CUDA. This resolves the cuda GLM prefill failed reproduction reported in #595 for the antirez/GLM-5.2-GGUF IQ2_XXS model with --ssd-streaming on DGX Spark / GB10.

This PR is based directly on current main and is independent of #592.

Root cause

The generic CUDA routed-MoE dispatcher accepted:

• Q4_K gate/up with Q4_K down
• IQ2_XXS gate/up with Q2_K down

Most routed layers in GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf use IQ2_XXS for gate, up, and down. The dispatcher rejected that combination before launching the routed projection, causing GLM prefill to fail.

Changes

• Add an IQ2_XXS down-projection CUDA kernel using the existing IQ2_XXS-to-f32 dot-product implementation.
• Dispatch IQ2_XXS gate/up/down layers through the existing f32 gate/up/SwiGLU path, the new IQ2_XXS down path, and the existing expert sum kernel.
• Preserve the existing Q4_K and IQ2_XXS-with-Q2_K paths unchanged.
• Add test-cuda-iq2-routed, a deterministic model-independent numerical test comparing every routed-MoE stage against the scalar CPU IQ2_XXS oracle.

Reproduction

Hardware: NVIDIA DGX Spark, GB10 (sm_121), CUDA backend.

./ds4 \
  --cuda \
  -m gguf/GLM-5.2-UD-IQ2_XXS_RoutedIQ2XXS_blk78Q2K.gguf \
  --ssd-streaming \
  --ctx 8192 \
  --temp 0 \
  --tokens 4 \
  -p 'Reply with exactly: OK'

Before this change:

ds4: prompt processing failed: cuda GLM prefill failed

After this change:

processing 23 input tokens: 23/23 (100.0%)
I need to reply
ds4: GLM prefill: 0.54 t/s, generation: 0.80 t/s

The process exited successfully after four non-EOS generation steps.

Numerical verification

make test-cuda-iq2-routed CUDA_ARCH=

gate: count=1024 worst_abs=3.33786e-06
up: count=1024 worst_abs=3.33786e-06
mid: count=1024 worst_abs=4.29153e-06
down: count=1024 worst_abs=8.39233e-05
out: count=512 worst_abs=8.39233e-05
test_cuda_iq2_routed PASS tokens=2 experts=2/3 dim=256

The test covers selected-expert routing, route weights, IQ2_XXS gate/up projections, SwiGLU, IQ2_XXS down projection, and expert summation.

Build verification

make ds4 ds4-server CUDA_ARCH=

Both CUDA binaries built successfully.

Fixes #595