feat: support TP>1 Domino rollout for DFlash V2

[jianuo-huang]

Summary

Extend the optimized Domino rollout from PR #36899 to tensor parallel (TP>1), including CUDA Graph capture and eager fallback paths.

• Compute first-token argmax from vocab-sharded LM-head logits with deterministic global tie handling.
• Add a full-vocabulary TP gather path for small logical tensors.
• Add a compact global candidate-pool path for larger batches/graphs: local top-k, TP all-gather, and candidate base-logit all-reduce.
• Validate LM-head and embedding shard metadata, padded vocab layout, TP rank, and unsupported added-vocab layouts.
• Keep B=1 on full gather and use compact candidate pooling for B>1 CUDA-graph buckets.

Correctness and validation

• SGLang nightly DLC unit test: 29 passed, 35 subtests passed.
• TP2 fake-shard tests compare both full-gather and compact paths against the TP1 full-vocabulary Domino oracle with exact token parity.
• Real 2-GPU TP2 Qwen3-8B service loaded the Domino checkpoint, captured decode graph buckets (bs=1 and bs=2), and completed serving successfully.
• Real 2-GPU TP2 Qwen3.6-27B service loaded both DFlash and Domino checkpoints, captured CUDA Graphs, and completed the full benchmark.

Qwen3.6-27B benchmark

Setting: TP2, BF16, FA3 full attention, block size 16, thinking enabled, max output 4096, greedy decoding, one measured run per cell. Workloads are GSM8K-128, HumanEval-164, and MT-Bench-80.

Workload	Concurrency	DFlash tok/s	Domino tok/s	Throughput gain	DFlash accept	Domino accept	Accept gain
GSM8K	1	340.25	437.40	+28.6%	6.777	9.287	+37.0%
GSM8K	8	1,855.85	2,314.51	+24.7%	6.860	9.304	+35.6%
GSM8K	32	3,094.09	3,794.29	+22.6%	6.840	9.286	+35.7%
HumanEval	1	363.60	399.53	+9.9%	6.422	7.407	+15.3%
HumanEval	8	2,048.48	2,206.18	+7.7%	6.451	7.433	+15.2%
HumanEval	32	3,452.99	3,733.11	+8.1%	6.496	7.422	+14.3%
MT-Bench	1	254.40	279.24	+9.8%	4.971	5.927	+19.2%
MT-Bench	8	1,442.02	1,517.15	+5.2%	5.027	5.966	+18.7%
MT-Bench	32	2,076.32	2,387.72	+15.0%	4.946	5.963	+20.5%

All 18 DFlash/Domino workload-concurrency cells completed successfully. Domino improves acceptance length and throughput in every measured cell.

Scope

This PR validates TP2 on the Qwen3 and Qwen3.6-27B model families, including eager and CUDA Graph paths. TP4/TP8, quantized LM-heads, and added-vocabulary shards are not claimed by this PR.

Related: #36899

---

CI States

Latest PR Test (Base): ✅ Run #34536644674
Latest PR Test (Extra): ❌ Run #34536644475
Latest PR Test (AMD ROCm 10): ❌ Run #34536644685