fix(voice): retain concrete language for turn detection

[Oxygen56]

Fixes #6530

Summary

• keep multi and auto language-detection modes from replacing the last concrete language
• share the same language-tracking logic across final and preflight transcripts
• add regression coverage for ambiguous multilingual results and later concrete-language updates

Root Cause

AudioRecognition stored every initial STT language value in _last_language. When a multilingual STT provider returned the mode value multi for an ambiguous utterance, the turn detector treated it as a concrete language code, could not find a matching threshold, and skipped end-of-turn prediction.

The STT event still retains the provider's original language value. Only the language tracked for language-specific turn-detection behavior ignores non-specific detection modes.

Verification

• uv run --no-sync pytest tests/test_audio_recognition_turn_detection.py::TestLanguageTracking --audio_eot -q -> 3 passed
• uv run --no-sync ruff check livekit-agents/livekit/agents/voice/audio_recognition.py tests/test_audio_recognition_turn_detection.py -> passed
• uv run --no-sync ruff format --check livekit-agents/livekit/agents/voice/audio_recognition.py tests/test_audio_recognition_turn_detection.py -> passed
• uv run --group typing mypy -p livekit.agents -> passed (205 source files)
• uv run --no-sync pytest --unit -q -> 1142 passed, 4 skipped before the repository's concurrent-test runner stopped with 9 event-loop errors on Python 3.13

Notes for Reviewers

• If no concrete language has been detected yet, multi leaves the tracked language unset and the existing endpointing fallback remains unchanged.
• No public API or STT event payload is changed.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[longcw]

could you sign the CLA before we merge?