perf: enable exact vocab-parallel greedy for target models

[sufubao]

Stack

This is the clean second layer on top of ModelTC/LightLLM#1517. It is opened in the fork so its diff contains only the target-model extension while ModelTC#1517 is under review. After ModelTC#1517 lands, this change can be proposed to upstream against main without repeating the shared reducer and draft-output contract.

Scope

• Add the opt-in LIGHTLLM_VOCAB_PARALLEL_GREEDY=1 target-model fast path.
• Enable it only when every request is unmodified greedy sampling.
• Retain dense logits for penalties, temperature changes, invalid-token masks, minimum-token EOS masking, constrained/diverse decoding, reward heads, and prompt-logprob capture.
• Keep CUDA Graph capture and replay on a stable dense-or-sparse output contract.
• Preserve sparse metadata through DP overlap row concatenation and backend sampling.
• Apply Gemma4 final-logit softcapping in FP32 on each local vocabulary shard before distributed reduction.

Verification

• 98 focused model-output, Triton, Gemma4, sampling, and speculative-decoding tests pass.
• Black, flake8, and git diff --check pass.
• The previously recorded full-stack Qwen3.5-27B TP=4/MTP=3 run improved median output throughput from 6791.56 to 7593.38 tok/s (+11.81%) and reduced median ITL from 8.589 to 7.562 ms (-11.96%). This measurement validates the combined draft + target stack; it does not isolate the marginal gain of this second layer.