[Bugfix][Spec Decode] Scope DSpark backend inheritance to DeepSeek V4 - #52809

Purpose

#52288 made an unspecified DSpark draft backend inherit the target's backend.
That fallback is required for DeepSeek V4 because its target and draft layers
share a KV-cache layout, but it is not valid for every DSpark architecture.

For example, a Kimi MLA target can use a Qwen3/SWA DSpark draft. Inheriting
FLASHINFER_MLA makes draft initialization fail because that backend does not
support sliding-window attention.

Keep the explicit draft override behavior, but inherit the target backend only
when the normalized DSpark draft has model_type == "deepseek_v4". Other
DSpark drafts leave the backend unset so their own attention kind participates
in normal backend auto-selection.

Why this is not a duplicate

This is a narrowly scoped follow-up to the regression introduced by #52288.
No open PR matched the backend-inheritance regression. Nearby #48381 addresses
the draft KV-cache dtype and #51042 addresses DSV4 SWA index width; neither
changes draft backend selection.

Tests

HF_HOME=/data/mgoin PYTHONPATH=/home/mgoin/code/vllm-dspark-backend-scope \
  /home/mgoin/code/vllm/.venv/bin/python -m pytest \
  tests/v1/spec_decode/test_dspark_utils.py -v
# 3 passed

/home/mgoin/code/vllm/.venv/bin/pre-commit run --files \
  vllm/v1/worker/gpu/spec_decode/dspark/utils.py \
  tests/v1/spec_decode/test_dspark_utils.py
# passed

/home/mgoin/code/vllm/.venv/bin/pre-commit run mypy-3.12 --hook-stage manual \
  --files vllm/v1/worker/gpu/spec_decode/dspark/utils.py \
  tests/v1/spec_decode/test_dspark_utils.py
# passed

The regression test covers DSV4 inheritance, non-DSV4 auto-selection, and an
explicit draft override.

Model validation

On 4x B300, main reproduced the non-DSV4 failure with a Kimi-K3 MLA target and
the Qwen3/SWA RedHatAI/Kimi-K3-speculator.dspark: the inherited
FLASHINFER_MLA backend was rejected because sliding-window attention is not
supported. With the independent-selection behavior restored, the target chose
FLASHINFER_MLA, the draft chose FLASHINFER, the engine initialized, and a
32-token chat request completed.

No output-accuracy evaluation was run because this changes startup-time backend
routing only. The DSV4 path retains the behavior validated in #52288.

AI assistance

This PR was produced with AI assistance. It is intentionally a draft; the human
submitter must review every changed line before marking it ready.