perf: avoid full-vocab gather for greedy sampling

[sufubao]

Summary

• add an opt-in vocab-parallel greedy path enabled by LIGHTLLM_VOCAB_PARALLEL_GREEDY=1
• reduce TP communication from full-vocabulary logits to three FP32 values per token and rank: local max, logsumexp, and argmax
• preserve the full-logits path for prompt logprobs and integrate the packed token/logprob result with normal and MTP post-processing
• fail explicitly for sampling modifiers and constrained logits that require mutable full-vocabulary logits

Performance

Qwen3.5-27B, 4x H100 80GB, TP=4, MTP=3, FP8, input/output=256/1024, concurrency=64, 128 requests. Both variants used the same server flags, including disabled FlashInfer allreduce, for a controlled comparison.

Variant	Run 1	Run 2	Median
Baseline	6405.09 tok/s	6578.85 tok/s	6491.97 tok/s
Vocab-parallel greedy	7245.27 tok/s	7539.74 tok/s	7392.51 tok/s

Median throughput improves by 13.87%. All 128 requests were valid with zero errors in every run.

Verification

• CUDA/Triton kernel and strict-greedy gate tests: 12 passed
• affected backend and MTP regression suites: 101 passed
• TP=4 numerical equivalence check with the real 151,936-token vocabulary at token counts 1, 7, 64, and 256: exact token IDs and logprobs within 2e-4
• formatting, Python compilation, and staged diff checks passed

Scope

This path is intentionally opt-in and supports only unmodified greedy requests: top-k 1, temperature 1, no penalties, invalid-token masks, minimum-new-token masking, or constrained logits.