[Bugfix][KV Connector][Mooncake] Apply multi-module MTP prefill backoff in P/D

[waizuichougou]

Purpose

Fix Model Runner V2 multi-module MTP with Mooncake P/D disaggregation.

The multi-module MTP scheduler exposes VllmConfig.num_prefill_lookahead_tokens
so the decoder can recompute the trailing lookahead window locally. The NIXL
connector already applies this backoff, but MooncakeConnector still transfers
the full prompt for dense MTP requests and does not truncate the producer-side
prefill request.

For example, with a prompt of 10 tokens and a three-token MTP lookahead,
Mooncake currently transfers 10 tokens while the decoder-side MTP path expects
the last 2 tokens to be recomputed locally. This can leave unverified MTP draft
KV in the P/D transfer state and make the decoder reuse KV that it does not
rebuild.

This change:

• applies the same prefill backoff as NIXL when calculating Mooncake's remote
  prefill token count;
• truncates the same trailing tokens on the producer before prefix-cache lookup;
• preserves the existing one-token Mamba backoff and leaves ordinary dense
  requests unchanged; and
• keeps the truncation idempotent across preemption and rescheduling.

This follows up on the multi-layer MTP P/D cache correction in #55055.

Test Plan

pytest -q \
  tests/v1/kv_connector/unit/test_mooncake_connector_hybrid_mamba.py
python -m py_compile \
  vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py \
  tests/v1/kv_connector/unit/test_mooncake_connector_hybrid_mamba.py

The regression tests cover the dense multi-module MTP P/D case, verify that
both sides use the same lookahead backoff, and confirm that ordinary dense
requests keep the existing full-prompt behavior. Existing hybrid Mamba tests
cover the unchanged one-token behavior.

Test Result

• 11 passed
• An end-to-end multi-module MTP P/D accuracy run has not been performed.

---

Essential Elements of an Effective PR Description Checklist

• The purpose and affected behavior are described.
• The test plan is included.
• Test results are included.
• No documentation or supported-model update is required.

[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.