azure tts: add streaming synthesis via Speech SDK WebSocket V2

[abhishekranjan-bluemachines]

Summary

• Add SynthesizeStream class to the Azure TTS plugin that uses the Azure Speech SDK's text streaming API (WebSocket V2 endpoint) to feed LLM tokens incrementally into the synthesizer
• Reduces time-to-first-audio-byte from ~550-1350ms (REST API, full sentence buffering) to ~100-200ms (streaming, progressive token feeding)
• Default behavior remains unchanged (use_streaming=False); opt-in via use_streaming=True

Details

• New stream() method returning SynthesizeStream — uses SpeechSynthesisRequest with TextStream input type on wss://{region}.tts.speech.microsoft.com/cognitiveservices/websocket/v2
• Two-task pipeline following the same pattern as the Google TTS plugin: sentence tokenization → sequential segment synthesis
• Bridges Azure SDK callbacks (background thread) to asyncio via loop.call_soon_threadsafe() + asyncio.Queue
• Pre-connects WebSocket per segment for lower TTFB
• Warns if SSML features (prosody, style, lexicon) are configured with streaming enabled, since text streaming does not support SSML
• Existing synthesize() / ChunkedStream REST API path fully preserved

Test plan

• Verify default use_streaming=False behavior is unchanged (REST API path)
• Test use_streaming=True with speech_key + speech_region auth
• Test use_streaming=True with speech_auth_token + speech_region auth
• Verify SSML warning is logged when streaming + prosody/style/lexicon configured
• Measure TTFB improvement in a cascaded STT→LLM→TTS voice agent pipeline
• Test cancellation and error handling (invalid key, network timeout)

[CLAassistant]

All committers have signed the CLA.

[abhishekranjan-bluemachines]

CI re-run: blockguard-tests (macos-latest) failed on a flaky timing-sensitive test (test_many_short_blocks) unrelated to this PR. Pushing this comment to retrigger CI.