feat(qwen3): add DFlash speculative decoding

[xiaguan]

Summary

• Add opt-in Qwen3-4B DFlash speculative decoding via --dflash-draft-model-path.
• Add native DFlash draft model loading/forward, verifier-span target execution, speculative KV lifecycle, and greedy acceptance logging with committed_tokens.
• Enable DFlash for all eligible active greedy requests when configured; unsupported combinations fail closed or use baseline execution.
• Harden multi-active DFlash with byte-budgeted side-state admission, transactional speculative KV rollback, active+pending prefill failure cleanup, and per-request prefill hidden capture.
• Split Qwen3 executor responsibilities into lifecycle / worker / DFlash lane / speculative execution modules, and document local/5090/vLLM comparison results.

Validation

• Commit hooks passed on pushed commits; targeted prek run --files passed for this hardening diff, including fmt and clippy.
• cargo fmt --all --check
• git diff --check
• cargo test --release -p openinfer-kv-cache --test lifecycle speculative -- --nocapture
• cargo test --release -p openinfer-qwen3-4b dflash_prefill --lib -- --nocapture
• cargo test --release -p openinfer-qwen3-4b admission_ --lib -- --nocapture
• cargo test --release -p openinfer-qwen3-4b dflash_ --lib -- --nocapture
• cargo test --release -p openinfer-qwen3-4b --lib -- --nocapture
• cargo build --release -p openinfer-server
• OPENINFER_TEST_MODEL_PATH=/data/models/Qwen3-4B OPENINFER_DFLASH_TEST_MODEL_PATH=/data/models/Qwen3-4B-DFlash-b16 cargo test --release -p openinfer-qwen3-4b --test hf_golden_gate dflash_speculative_verify_matches_hf_argmax_regret_gate -- --nocapture
• Toxic review re-check after the multi-active hardening returned Ready for review with no P0/P1/P2 findings.
• Local 5070 Ti PR-head vllm bench serve, greedy /v1/completions:
  • Spec-Bench c1: 89.72 -> 149.32 tok/s (1.66x)
  • Spec-Bench c4: 303.92 -> 330.42 tok/s (1.09x)
  • Random 1024/128 c1: 86.61 -> 136.09 tok/s (1.57x)
  • Random 1024/128 c4: 270.93 -> 349.50 tok/s (1.29x)
• Post-hardening local Spec-Bench c4 smoke completed 12/12 at 368.71 tok/s and logged four concurrent DFlash request IDs in one wave.
• 5090 real-weight gate passed with /data/Qwen3-4B and /data/Qwen3-4B-DFlash-b16.
• 5090 reference measurements are documented: OpenInfer Spec-Bench c1 167.34 -> 251.48 tok/s; upstream vLLM 0.22.1 DFlash reaches 289.57 tok/s on the same Spec-Bench with native acceptance metrics.

Draft State

• Draft PR opened early to avoid branch collision.
• Latest code removes the single-active-request DFlash restriction; multi-active greedy requests are now eligible together.
• Known performance shape: target verify is batched, but DFlash draft is still per-request serial and not CUDA-graph captured.
• Main remaining performance work before calling this polished is bs=1 draft-path optimization against the 5090 vLLM DFlash reference, then replacing serial draft with graph-captured or batched/fused draft execution.

[wjinxu]

Heads-up + I'd like to build on this rather than ship a parallel stack: I've been prototyping n-gram (prompt-lookup) speculative decoding in #349. n-gram is just a CPU draft source that produces the same kind of verify span, so it should be able to reuse your verify + greedy-acceptance + schedule/apply/revert_speculative KV path as a zero-cost second method (no draft model, no draft KV, no hidden capture).

I'll hold the detailed integration until your PR settles so I'm adapting to the final shape, not a moving target.