Add NeMo backend for Parakeet support in server whisper

[basnijholt]

Summary

• add a new nemo ASR backend for agent-cli server whisper with subprocess lifecycle/TTL compatibility
• add support for NVIDIA Parakeet model shorthand parakeet-tdt-0.6b-v2 (resolved to nvidia/parakeet-tdt-0.6b-v2)
• wire nemo into backend selection/factory and --download-only
• update server whisper CLI help/docs with NeMo + Parakeet examples
• make --backend auto switch to nemo for Parakeet models and return a clear error if Parakeet is requested with non-NeMo backends
• add tests for NeMo model alias/download helper and CLI help visibility

Validation

• pre-commit hooks passed on commit (ruff, mypy, jscpd, pylint duplicate-code)
• pytest -q tests/test_nemo_backend.py tests/test_api_integration.py::test_server_whisper_command_in_cli tests/test_server_whisper.py

[basnijholt]

Follow-up updates pushed:

• Added nemo-whisper optional dependency in pyproject.toml and synced:
  • agent_cli/_extras.json
  • agent_cli/_requirements/*.txt (including new nemo-whisper.txt)
  • uv.lock
• Restored strict command gating on server whisper:
  • @requires_extras("server", "faster-whisper|mlx-whisper|whisper-transformers|nemo-whisper", "wyoming")

Why the decorator was briefly removed:

• Initially there was no NeMo extra to satisfy requires_extras for --backend nemo, so that check would fail before backend selection.
• This is now fixed correctly with a dedicated nemo-whisper extra.

Live tests run (real commands):

1. Generated audio via CLI:
   • agent-cli speak 'This is a live test for parakeet backend validation.' --tts-provider wyoming --tts-wyoming-ip localhost --tts-wyoming-port 10200 --save-file /tmp/live_speak.wav --output-device-name 'Dummy Output'
   • Result: /tmp/live_speak.wav created successfully.
2. Started dedicated NeMo server on alternate port:
   • agent-cli server whisper --backend nemo --model parakeet-tdt-0.6b-v2 --host 127.0.0.1 --port 11301 --wyoming-port 11300 --no-wyoming
3. Transcribed generated audio through Parakeet endpoint:
   • curl -X POST http://127.0.0.1:11301/v1/audio/transcriptions -F file=@/tmp/live_speak.wav -F model=parakeet-tdt-0.6b-v2
   • Response text: This is a live test for parakeet back-in validation.
   • First request (model load): ~22.8s
   • Second request (model warm): ~0.57s
4. Verified CLI path with file input + OpenAI-compatible ASR provider:
   • agent-cli transcribe --from-file /tmp/live_speak.wav --asr-provider openai --asr-openai-base-url http://127.0.0.1:11301/v1 --asr-openai-model parakeet-tdt-0.6b-v2 --json
   • Returned same transcript successfully.

Validation:

• pre-commit hooks passed on latest commit.
• Targeted tests passed:
  • tests/test_nemo_backend.py
  • tests/test_api_integration.py::test_server_whisper_command_in_cli
  • tests/test_server_whisper.py

[basnijholt]

Blocked by upstream NeMo dependency pinning discussion/update: NVIDIA-NeMo/NeMo#15438