[Bugfix][LoRA] Fix Qwen3-ASR audio tower QKV adapters

[ksellesk]

Purpose

Follow-up to #37247.

Qwen3-ASR's audio tower exposes attention Q/K/V as a fused
QKVParallelLinear named qkv, while PEFT checkpoints store separate
q_proj, k_proj, and v_proj LoRA weights. The model currently only
declares the qkv_proj packed mapping used by the language model, so the
LoRA manager receives an empty packed-module list for
audio_tower.layers.*.self_attn.qkv and ignores those adapters.

This PR adds the missing qkv packed mapping and a regression test that
checks the Qwen3-ASR audio QKV layer is wrapped with
MergedQKVParallelLinearWithLoRA.

User impact: all-linear Qwen3-ASR PEFT adapters can apply their audio tower
Q/K/V LoRAs instead of silently skipping them.

Duplicate check: searches for open Qwen3-ASR audio tower/QKV LoRA issues and
PRs found no existing fix. #49525 addresses text-only multimodal mapper
fallbacks, and #45944 targets Qwen2-Audio, so neither overlaps this change.

Test Plan

.venv/bin/python -m pytest \
  tests/lora/test_lora_manager.py::test_wrap_qwen3_asr_audio_qkv -q

End-to-end validation uses Qwen3-ASR-1.7B with an all-linear rank-64 PEFT
adapter and --enable-tower-connector-lora, then calls
/v1/audio/transcriptions with a LibriSpeech dev-clean sample.

Test Result

Static checks on the latest main checkout:

.venv/bin/python -m py_compile ...  PASS
git diff --check                    PASS

End-to-end validation on vLLM 0.26.0:

• before the mapping change, all 24 audio encoder QKV layers logged
  could not be wrapped ... It will be ignored;
• after the change, all 24 warnings disappeared and the adapter loaded;
• /v1/audio/transcriptions returned HTTP 200;
• LibriSpeech sample 652-129742-0020 exactly matched the reference:
  STRAIN AND BOTTLE AND PUT IN ICE BOX SHAKE BEFORE USING EACH TIME.

The targeted pytest was not run in this checkout because installing the
current main-branch PyTorch 2.13/CUDA 12.9 test environment was stopped before
completion. The PR is opened as a draft so CI and reviewer feedback can cover
that remaining check.

No documentation update is needed because this corrects the existing
Qwen3-ASR tower LoRA behavior.

Developed with assistance from OpenAI Codex. I reviewed every changed line and
the validation evidence.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this 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.
• Documentation impact considered; no update is required for this behavior fix.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment /ci run whenever CI signals are needed.

Once the PR is approved or has the ready label, the PR author can also use /ci run or /ci retry. New commits do not start CI automatically.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀