[Spec Decode][DCP] Populate draft-local sequence lengths

[GirasoleY]

Summary

• Recompute rank-local DCP sequence lengths from the draft model's current global sequence lengths before every draft metadata build.
• Use the finalized BlockTables CP size, rank, and interleave instead of copying parallel state into the speculator.
• Share the logic across EAGLE, MTP, and DFlash and remove DFlash's duplicate preparation.
• Add an EAGLE regression test covering padded DCP-local metadata.

Motivation

Draft attention owns separate persistent metadata from the target. EAGLE and autoregressive MTP currently call the shared metadata builder without dcp_local_seq_lens; MLA metadata then receives None under DCP. Reusing target-local lengths is not sufficient because draft steps advance or rewind input_buffers.seq_lens after sampling and rejection.

This completes the draft-side metadata plumbing needed by DCP-capable attention backends such as TokenSpeed MLA (#48180) and FlashInfer MLA (#54012).

Tests

Passed locally:

• uvx --offline ruff check on changed files
• uvx --offline ruff format --check on changed files
• python3 -m py_compile on changed files
• git diff --check

Focused pytest and multi-GPU EAGLE+DCP evaluation are pending because this local worktree does not contain the vLLM runtime/test dependencies. This PR is intentionally opened as a draft.

AI assistance disclosure

AI assistance was used to trace the EAGLE/DFlash metadata paths, adapt the k3-dev behavior to current upstream BlockTables state, write the initial patch and regression test, and run static validation. The submitter will review every changed line and run focused/runtime tests before requesting final review.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[GirasoleY]

Superseded by #54277, which now contains this generic draft-side DCP metadata commit together with the FlashInfer MLA DSpark+DCP backend changes.