[Docs] Add an audio tagging, quality-metrics, normalization, and LLM PnC guide #2123

Context

Child of #2118. #1863 added a substantial audio-tagging pipeline, but its only Fern change was to the speaker-separation page. The main guidance remains in tutorials/audio/tagging/README.md.

APIs to cover

• PNCwithvLLMInferenceStage / CleanLLMOutputStage
• Reusable VLLMInference lifecycle
• WER, CER, SQUIM, and bandwidth stages
• PrepareModuleSegmentsStage
• ITN, Arabic diacritic removal, and Chinese conversion
• ASR and TTS pipeline configurations

Requirements

• Explain stage ordering, expected manifest schema, columns added/removed, and model/dependency setup.
• Document first-pass/second-pass PnC, CER validation, fallback behavior, and operational limits.
• Include an end-to-end runnable example and a configuration reference for both ASR and TTS paths.
• Cross-link speaker diarization/separation and audio quality filtering.

Acceptance criteria

• User-facing Fern documentation exists beyond the GitHub tutorial README
• Pipeline diagram and stage table match the merged implementation
• Dependencies/extras and vLLM constraints are explicit
• Inputs, outputs, defaults, and failure behavior are documented
• Page is in audio navigation and cross-linked from text integration/quality pages
• Fern checks pass

Related PR: #1863

Supplemental code-freeze attribution

#1679 introduced the foundational audio tagging pipeline, including resampling, PyAnnote diarization, WhisperX VAD, NeMo ASR alignment, segment merging, YAML configuration, and the initial tutorial. Treat #1679 and #1863 as one implementation history and document the final API on main.

Related PRs: #1679, #1863