(retriever) ASR refactor

[charlesbluca]

Description

Refactors ASR so all “chunk rows → transcript rows” logic goes through asr_chunks_to_text, with ASRActor as a thin wrapper.

What changed

• asr_chunks_to_text(batch_df, model=..., client=..., asr_params=...)
  Single batch entry point for ASR. ASRActor only constructs model/client from ASRParams and delegates here.

• Injectable model / client
  Inprocess and the GPU pool can pass a ParakeetCTC1B1ASR (or remote client) so the same code path runs inside and outside Ray map_batches.

• Long audio (local Parakeet)
  ParakeetCTC1B1ASR splits inputs that exceed the model length budget and concatenates transcripts.

• Media probing
  media_interface: more robust ffprobe handling when duration or bit_rate is missing (e.g. VBR / bad probes).

• Inprocess
  _load_doc_to_df / _iter_doc_chunks unify loading for pdf / html / image / audio / txt in the ingest loop.

• API
  Removed apply_asr_to_df; use asr_chunks_to_text directly (tests updated).

Why

• One implementation for Ray batch, inprocess, and audio CLI—easier to fix and extend.
• Injected model supports GPU pool / avoids duplicate model setup where the caller already holds the model.
• Preserves main behavior for remote segment_audio while keeping the refactored structure.

Testing

• pytest nemo_retriever/tests/test_asr_actor.py

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• If adjusting docker-compose.yaml environment variables have you ensured those are mimicked in the Helm values.yaml file.