perf: avoid full-vocab all-gather for draft greedy sampling

[sufubao]

Motivation

Draft/MTP proposal generation only consumes greedy argmax tokens, so gathering every vocabulary logit onto every TP rank wastes communication and memory. #1512 addresses the same bottleneck with gathered top-K candidates; this version keeps both token selection and selected-token probability exact over the complete vocabulary.

This is the first half of the previously broader PR. It deliberately covers draft models only. The target-model opt-in, sampling eligibility checks, dense fallback, Gemma4 handling, and generic backend integration are isolated in a follow-up branch.

Design

• Draft models are detected through the existing is_mtp_draft_model model invariant and use the sparse path automatically.
• Each TP rank computes its local maximum, local logsumexp, and local argmax, then performs one fixed-size all-gather of three FP32 words per token.
• Global ties deterministically select the smallest token ID. Argmax IDs are bit-preserved through the FP32 communication payload.
• ModelOutput.logits remains a real logit tensor. Global IDs and full-vocabulary normalization are carried explicitly as logits_token_ids and logits_logsumexp.
• Draft argmax and confidence paths reconstruct global token IDs and exact probabilities as exp(selected_logit - full_vocab_logsumexp).
• CUDA Graph padding/unpadding, Eagle accepted-row selection, dynamic MTP, and DSpark confidence sampling preserve the explicit metadata.
• Target-model sampling is unchanged in this PR; no feature flag or sampling fallback is introduced here.

Verification

• 26 focused model-output, Triton reducer, Eagle, and DFlash tests pass.
• The reducer tests compare token IDs, selected logits, logsumexp, and logprobs against a materialized full vocabulary, including local and cross-rank ties.
• Black, flake8, and git diff --check pass.

Follow-up

The isolated target-model extension is available as a stacked comparison. It will be proposed to upstream after this draft-only contract lands, so reviewers do not need to review the shared kernel and output plumbing twice.

Related: #1511, #1512.