fix azure multilingual voice language SSML

[mustafatarim]

Summary

Fix Azure TTS SSML generation for multilingual voices when the requested text language differs from the voice's primary locale.

Azure multilingual voices such as en-US-JennyMultilingualNeural can synthesize other languages, but the generated SSML currently places the requested language only on the root <speak> element while the voice itself remains cross-locale. For cross-locale multilingual synthesis, this change wraps the spoken content in <lang xml:lang="..."> and keeps the root language aligned with the voice locale.

Changes

• Detect the primary locale from Azure voice names.
• Wrap content with <lang xml:lang="..."> only for MultilingualNeural voices when language differs from the voice locale.
• Preserve existing behavior when no explicit language is provided, when the voice and language locales match, or when the voice is not multilingual.
• Add Azure TTS SSML unit tests for cross-locale multilingual and non-wrapped cases.

Validation

• uv run pytest tests/test_plugin_azure_tts.py
• uv run ruff check livekit-plugins/livekit-plugins-azure/livekit/plugins/azure/tts.py tests/test_plugin_azure_tts.py
• uv run ruff format --check livekit-plugins/livekit-plugins-azure/livekit/plugins/azure/tts.py tests/test_plugin_azure_tts.py

[CLAassistant]

All committers have signed the CLA.