[Diffusion] Fuse Qwen-Image text QKV projections

[BBuf]

Motivation

Qwen-Image projects the short text stream with three separate tensor-parallel Q/K/V GEMMs in every transformer block. vLLM-Omni uses one merged QKV projection. A torch-profiler comparison on 2x H200 showed that merging only the short text branch removes two small split-K launches per block without exposing the large image stream to packed-view materialization costs.

This PR is intentionally separate from the bitwise attention packing optimization because a merged BF16 GEMM changes reduction association.

Modifications

• Use MergedColumnParallelLinear for the unquantized Qwen-Image added/text QKV projection.
• Add checkpoint mappings that merge Diffusers add_q_proj, add_k_proj, and add_v_proj weights in Q/K/V order.
• Preserve the existing unfused layout for other quantization backends; Nunchaku remains on its existing fused path.
• Add CPU-only wiring and loader-mapping tests.

Accuracy Tests

Validated on the latest main used for this PR (b647ae8) and 2x NVIDIA H200:

• python3 test/registered/unit/models/test_qwen_image_fused_qkv.py: 3 passed.
• Five repeated candidate generations were deterministic and byte-identical to each other.
• Versus the separate-GEMM BF16 baseline: pixel MAE 1.393/255, PSNR 40.55 dB. The difference is expected from GEMM reduction-order reassociation; no model weights or mathematical operation change.

This remains a Draft so maintainers can decide whether that BF16 numerical delta is acceptable for the default speed path.

Speed Tests and Profiling

Environment: 2x H200, TP=2, Qwen/Qwen-Image-2512, 1024x1024, 50 steps, guidance=1, BCG enabled, torch.compile disabled. The incremental A/B used the same CustomAllReduceV2 stack on both sides.

• Steady denoise baseline: 78.5-78.8 ms/step.
• Text-QKV candidate: 78.0-78.5 ms/step (78.2 ms median), about 0.4 ms/step / 0.5% faster.
• Torch trace: removed 1,212 split-K GEMMs (7.606 ms) and 1,212 split-K reductions (3.108 ms); added 403 merged GEMMs (4.797 ms) and the expected packed-slice copies (3.161 ms).
• Isolated projection microbench, CUDA graph: text seq=64 separate 31 us vs merged 13 us. The large image branch is deliberately left unfused because full-model A/B showed it regressed denoise to about 82 ms/step.

Checklist

• Format your code according to the contribution guide.
• Add unit tests according to the contribution guide.
• Update documentation according to the contribution guide (no user-facing API or flag change).
• Provide accuracy and speed benchmark results.
• Follow the SGLang code style guidance.

Background benchmark investigation: BBuf/how-to-optim-algorithm-in-cuda#21

---

CI States

Latest PR Test (Base): ❌ Run #33069116662
Latest PR Test (Extra): ❌ Run #33069116477
Latest PR Test (AMD ROCm 7.2): ❌ Run #33069116717

[BBuf]

Consolidated into #36680 as requested. The full text-QKV commit and its tests are now preserved there.