feat: add optimized Domino rollout to DFlash V2

[jianuo-huang]

Summary

This PR adds optimized Domino projector support to DFlash V2 for the public
Qwen3-8B-Domino-b16 checkpoint. It is based on the latest upstream
main and supersedes the earlier Domino PRs #31328 and #36860.

The target verification, continuous-prefix acceptance, bonus-token handling,
and KV-cache commit paths remain unchanged.

Included implementation

• Parse, validate, instantiate, and load the bias-free prefix GRU and the
  two-layer SiLU correction projector.
• Generate the 15-token Domino proposal chain for block size 16, feeding each
  realized proposal back into the GRU.
• Precompute all base LM-head logits in one step-major projection.
• Capture the complete rollout in the DFlash draft CUDA Graph, with eager
  fallback when graphs are unavailable.
• Use ATen GRUCell for one-token feedback updates to avoid repeated cuDNN
  weight packing.
• Add a request-local, block-shared K=2048 candidate vocabulary pool for the
  correction head. K=0 remains the full-vocabulary reference path.

Scope

• BF16, CUDA, TP=1, Qwen3-8B, DFlash block size 16.
• Greedy decoding path; target verification remains full-vocabulary and is the
  correctness boundary.
• The candidate pool is a draft-side approximation; it does not alter target
  verification semantics.
• Non-greedy sampling and tensor-parallel Domino are covered by follow-up work
  in PR feat: support TP>1 Domino rollout for DFlash V2 #37069.

Validation

Unit and registered tests

Focused tests:

PYTHONPATH=python python3 -m pytest -q \
  test/registered/unit/spec/test_dflash_domino.py

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=python3 -m pytest -q -s \
  test/registered/spec/dflash/test_dflash_domino.py

The focused suite covers config parsing, strict Domino projector weight loading,
shape validation, full-chain rollout parity, candidate-pool boundaries,
batched-vs-single-row parity, eager fallback, and CUDA Graph replay.

Serving validation setting

We validated the Qwen3-8B DFlash/Domino serving path with the following fixed setting:

• Target: Qwen/Qwen3-8B
• Drafts: Qwen3-8B-DFlash-b16 and Qwen3-8B-Domino-b16
• CUDA, BF16, TP=1, FA3 attention backend
• DFlash block size: 16 (15 proposed tokens)
• Domino candidate pool: K=2048
• Greedy sampling: temperature=0, top_p=1, top_k=1
• Qwen3 tokenizer chat template, thinking disabled
• Workloads: GSM8K-128, HumanEval-164, MT-Bench-80
• Concurrency: 1, 8, 32; maximum generation length: 2048
• Server: max_running_requests=32, cuda_graph_max_bs=32,
  mem_fraction_static=0.70, page_size=1
• Warmup: 3 requests, warmup maximum generation 1024, first request skipped
• One measured run per configuration, timeout 3600 seconds

The serving validation completed successfully for both DFlash and Domino on this setting.
Numerical throughput and acceptance-length comparisons are reported separately in PR #37069.

Changes

This branch is rebased directly onto the current upstream main; no unrelated
benchmark or training scripts are included.

---

CI States

Latest PR Test (Base): ❌ Run #34294845632
Latest PR Test (Extra): ❌ Run #34294845436
Latest PR Test (AMD ROCm 10): ❌ Run #34294845773