perf: share vocab-parallel top-k logits for target and draft

[sufubao]

Summary

Replace the exact vocab-parallel greedy output with a shared top-k candidate path, following the approach in #1512.

• take the local top-k logits on each TP rank
• pack FP32 logits and INT32 global token IDs into one fixed-shape all-gather
• reuse the same sparse output representation for draft and target models
• keep draft models on the sparse path by default
• enable target-model sparse output only for unmodified greedy batches; penalties, EOS masking, invalid-token masking, constrained sampling, and other incompatible cases fall back to dense logits
• remove the full-vocabulary logits_logsumexp bookkeeping and the dedicated exact-greedy kernels
• preserve sparse metadata through CUDA Graph padding, microbatch overlap, dynamic MTP, Gemma4 logit softcap, and Qwen3-DSpark confidence handling

Greedy token selection remains exact because every TP shard contributes its local maximum among the candidates. Probabilities returned from sparse logits are intentionally approximate candidate-set probabilities.

Configuration

• LIGHTLLM_VOCAB_PARALLEL_TOPK=1: enable the target-model sparse path
• LIGHTLLM_VOCAB_PARALLEL_TOPK_SIZE: local candidate count per TP rank, default 128

Draft models always use the shared top-k path.

Validation

• focused top-k/model-output tests: 28 passed
• model-infer and Qwen3-DSpark regression tests: 129 passed
• Black, flake8, compileall, and git diff --check: passed