feat(services/qwen): add Qwen3STTService for local speech-to-text

[RoKenobi]

Summary

Adds Qwen3STTService — a local speech-to-text service backed by the Qwen3-ASR model family via the qwen-asr package.

• Extends SegmentedSTTService, following FunASRSTTService / WhisperSTTService / MoonshineSTTService
• Three model sizes: Qwen/Qwen3-ASR-0.6B, Qwen/Qwen3-ASR-1.7B (default), Qwen/Qwen3-ASR-8B
• Multilingual: English, Chinese, Spanish, French, German, Japanese, Korean, Portuguese, Russian, Arabic, Italian, Hindi
• Runs on CUDA via torch.bfloat16; transcription dispatched with asyncio.to_thread
• Unit tests with a mocked qwen-asr, following the FunASR/Moonshine test patterns

Replaces #5415, which was opened against a three-week-old main and had drifted far enough that rebasing in place was messier than starting clean.

Packaging: conflicting extras

qwen-asr depends on accelerate==1.12.0, while the moondream extra constrains the same package to ~=1.10.0. uv resolves a single universal lockfile across all extras, so the two pins are unsatisfiable together and uv lock fails outright — regardless of which extras a given install actually selects. That is what broke the docs build on #5415.

Since nobody runs a local Qwen3-ASR model and Moondream vision in one environment, the two are declared conflicting:

conflicts = [
    [
        { extra = "moondream" },
        { extra = "qwen-asr" },
    ],
]

uv then resolves them in separate forks, and activating both together becomes an explicit install-time error rather than a resolution failure for everyone.

qwen-asr is also excluded in docs/api/install-deps.sh alongside the other heavy ML extras (it pulls in torch), and qwen_asr is added to autodoc_mock_imports so the service's API page still generates.

Usage

from pipecat.services.qwen.stt import Qwen3STTService, Model
from pipecat.transcriptions.language import Language

stt = Qwen3STTService(
    model=Model.ASR_1_7B,   # or "Qwen/Qwen3-ASR-0.6B", "Qwen/Qwen3-ASR-8B"
    device="cuda:0",
    language=Language.EN,
)

Install: uv add "pipecat-ai[qwen-asr]"

Not done here

uv.lock is not regenerated — this branch carries main's lockfile, which predates the qwen-asr extra. Read the Docs re-resolves during uv sync, so the docs build exercises the conflicts change, but uv lock has not been run against it directly.

Test plan

• Unit tests: defaults, settings overrides, string model ID, language mapping, run_stt transcription, missing-model error, inference-exception error, empty-transcript skip
• Validated in production prior to this PR in a pipecat-based voice server (Qwen3-ASR-1.7B on CUDA)
• uv lock regenerated and committed

Notes

• No changes to the existing qwen/llm.py or qwen/__init__.py; the qwen = [] extra (DashScope LLM) is untouched and qwen-asr is separate
• Each service instance loads its own copy of the model into VRAM. The class docstring documents the subclass pattern for sharing one Qwen3ASRModel across pipeline sessions.