feat(aic): update ai-coustics integration to SDK 0.21 (VAD probability + Tyto)

[gkmngrgn]

Summary

Updates the ai-coustics integration to the newly released SDK 0.21 (Python wrapper aic-sdk 2.5.0, native 0.21.0). Two independent commits.

Commit 1 — VAD probability + new Quail models

• Bump the aic-sdk pin ~=2.3.0 → ~=2.5.0. The old pin resolved to >=2.3.0,<2.4.0, which blocked 2.4/2.5 — nobody on Pipecat could install the newest aic-sdk.
• AICQuailVADAnalyzer.voice_confidence() now returns the model's continuous raw probability (VadContext.raw_vad_probability(), new in 0.21) instead of the binary is_speech_detected() flag, so Pipecat's own VADParams (confidence/start_secs/stop_secs) performs the speech gating — the intended VADAnalyzer contract.
• Removed the now-redundant SDK-side sensitivity/speech_hold_duration/minimum_speech_duration params (they only affected the unused post-processed path; the analyzer is unreleased, so no deprecation needed).
• Default VAD model → quail-vf-vad-2.0-s-16khz; example enhancement model → quail-vf-2.2-l-16khz.

Commit 2 — Tyto audio-quality analyzer

• New AICTytoAnalyzer FrameProcessor backed by the ai-coustics Tyto model (aic-sdk 2.4.0+). Passively taps input audio, buffers into the SDK Collector, and runs Analyzer.analyze_buffered() off the event loop on a configurable interval.
• Emits a new AICAudioQualityMetricsData — seven 0.0–1.0 scores (risk_score, speaker_reverb, speaker_loudness, interfering_speech, media_speech, noise, packet_loss) predicting downstream STT/VAD/turn-taking degradation — via a MetricsFrame and an on_audio_analysis event.
• RTVIObserver forwards the scores to clients under an audio_quality metrics key.
• Adds a voice-aicoustics-audio-quality example and unit tests.

Test plan

• uv run ruff check / uv run ruff format --check — clean.
• uv run pytest for the aic suites + RTVI + deprecation markers — 185 passed.
• uv run towncrier build --draft — both changelog entries render.

Remaining before ready (why this is a draft)

• Changelog fragment names: currently orphan +… files (render with an empty (PR )). Rename to <PR#>.changed.md / <PR#>.added.md once this PR number is known.
• Tyto streaming smoke test with a real AIC_SDK_LICENSE: unit tests mock the SDK. Need to confirm on real hardware that Collector.buffer() accepts variable-size chunks under allow_variable_frames=True, and that the input sample rate is resampled to the model rate internally. Run examples/voice/voice-aicoustics-audio-quality.py.
• Confirm whether Tyto warrants its own PR vs. shipping alongside the VAD work.