[Bugfix][Qwen3.5] Keep the MTP drafter stage-local under pipeline parallelism

[Peuqui]

Purpose

Qwen3.5-family targets (Qwen3.5, Qwen3.6, Qwen3.8 dense and MoE) with
--speculative-config method=mtp and --pipeline-parallel-size greater than one
do not boot. There are three stacked causes, each hiding the next:

1. Qwen3_5MTP does not implement SupportsPP. The draft model config is
   verified against the parallel config, so the engine refuses to start:

   NotImplementedError: Pipeline parallelism is not supported for this model.
   Supported models implement the `SupportsPP` interface.
   

2. Qwen3_5MultiTokenPredictor.forward branches on get_pp_group().is_first_rank,
   which is the TARGET model's pipeline position. The drafter only runs on the
   last rank (execute_model returns the IntermediateTensors on every earlier rank
   before speculation), so it took the "receive from the previous stage" path and
   asserted on intermediate tensors that nobody sends. This is the same defect
   [Bugfix][Qwen4Exp] Keep the MTP drafter stage-local under pipeline parallelism #573 fixed for Qwen4Exp.

3. With the default VLLM_QWEN35_MTP_SHARE_IO_WEIGHTS=1 the drafter builds its
   embedding as a PPMissingLayer and expects the loader to share the target's.
   Both load_eagle_model (V2) and _maybe_share_embeddings (V1) deliberately
   skip embedding sharing under PP, because the target's embedding only exists on
   the first rank. The placeholder was therefore never replaced, and the first
   compile on the last rank failed with

   torch._dynamo.exc.Unsupported: Assertion failed on symbolic shapes
     assert hidden_states.shape[-1] == inputs_embeds.shape[-1]
   

Changes:

• Qwen3_5MTP implements SupportsPP and forwards
  make_empty_intermediate_tensors from the predictor, as Qwen4ExpMTP does.
• The predictor's forward drops both pipeline branches: it always embeds locally
  and never returns IntermediateTensors for a next stage it does not have.
• Under PP the drafter owns its embedding (loaded from the checkpoint's
  embed_tokens) instead of waiting for a share that the loaders skip. With a
  single pipeline rank nothing changes: sharing stays as before. The LM head is
  still shared, since the target's head lives on the drafter's rank.

Single-rank behaviour is unchanged: there is_first_rank and is_last_rank are
both True, the removed branches were dead, and world_size == 1 keeps the shared
embedding.

Not a duplicate: no open PR here or in vllm-project/vllm touches Qwen3_5MTP under pipeline parallelism; #573 fixed only the Qwen4Exp drafter.

Test Plan

.venv/bin/python -m pytest tests/models/qwen3_5/test_mtp_stage_local.py -v
pre-commit run --files <changed files>
pre-commit run mypy-3.10 --hook-stage manual --files <changed files>

End to end on 2x Quadro RTX 8000 (TP1 x PP2), RadixArk/Qwen3.8-27B-NVFP4,
--speculative-config '{"method":"mtp","num_speculative_tokens":3,"draft_sample_method":"greedy"}',
fp16, V2 model runner, 400 tokens at temperature 0, five repetitions, plus the same
prompt on TP2 with a DFlash2 draft as a text reference.

Test Result

• New tests: 3 passed. With the model file reverted to main: 3 failed
  (AssertionError on the intermediate tensors, missing SupportsPP).
• main, TP1 x PP2: the engine refuses to start with the NotImplementedError above.
• With causes 1 and 2 fixed but the embedding still shared: boot fails on the last
  rank with the dynamo assertion quoted above.
• This PR, TP1 x PP2: boots in 105 s and serves. 61.05 tok/s median (61.00 to
  61.10), acceptance length 2.963. The 400-token greedy text is byte-identical to
  the TP2 reference (sha256 prefix 0106659946c064b1), so the drafter changes
  acceptance, never the output.

AI assistance (Claude) was used to write this change and this description. I
reviewed every changed line and ran the tests and measurements above.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.