Draft: PnC with LLM for audio pipeline

[sushmitha-deva-09]

Description

Adds Punctuation and Capitalization (PnC) with LLM inference to the audio tagging pipeline. This enables using vLLM-backed language models (e.g. Qwen/Qwen2.5-1.5B-Instruct) to generate punctuated text from ASR output, with validation and fallback logic.

Key components

• PNCwithvLLMInferenceStage — Batched GPU inference stage that generates punctuated/capitalised text using vLLM. Supports both segment-level and top-level text processing.
• CleanLLMOutputStage — Post-processing stage that validates LLM output against original ASR text using CER, flags entries exceeding the threshold with pnc_fallback=True, and optionally updates word-level alignment timestamps.
• VLLMBase model interface — Shared vLLM engine wrapper with prompt templating, tokenizer management, and GPU lifecycle management.

Usage

 stages:
  - _target_: nemo_curator.stages.audio.tagging.text.pnc.PNCwithvLLMInferenceStage
    name: "PNCwithvLLM"
    model_name: "Qwen/Qwen2.5-1.5B-Instruct"
    text_key: "text"
    generation_field: "text_pnc"
    batch_size: 64
    resources:
      gpus: 1.0

  - _target_: nemo_curator.stages.audio.tagging.text.pnc.CleanLLMOutputStage
    name: "CleanLLMOutput"
    generation_field: "text_pnc"
    asr_pred_text_key: "text"
    cer_threshold: 2.0

Checklist

• I am familiar with the Contributing Guide.
• New or Existing tests cover these changes.
• The documentation is up to date with these changes.

[sushmitha-deva-09]

/ok to test e95efb7