feat: add proxy RAG server (#84)

Restore portaudio instructions for Linux (#85)

Change from **stream_kwargs to typed StreamConfig (#83)

feat(cli): add -h alias for --help (#77)

Add `--tts-openai-base-url` flag (#75)

Add custom ASR provider support (#73)

## Summary
- Add an OpenAI-compatible NVIDIA ASR server (Canary/Parakeet) with docs and locked deps for reproducible `uv run` usage.
- Let OpenAI ASR requests target a custom base URL/prompt (for self-hosted Whisper-style endpoints) and relax API key requirements when a custom endpoint is used.
- Fix ffmpeg resampling in the NVIDIA server so regular WAV/MP3 uploads are decoded correctly, and update tests/config for the new ASR options.

## Key Changes
- New `scripts/nvidia-asr-server/` with FastAPI server, README, `pyproject.toml`, `uv.lock`, and Nix shell for running NVIDIA ASR models locally.
- ASR OpenAI config/options now include `asr_openai_base_url` and `asr_openai_prompt`; ASR calls use a dummy key when hitting custom endpoints.
- Transcribe tests updated to pass the new ASR options; misc cleanup + gitignore removal so the server assets are committed.
- ffmpeg input flags removed so uploaded audio formats are auto-detected before resampling to 16 kHz mono.

## Usage Example
```bash
# Start NVIDIA ASR server (OpenAI-compatible)
cd scripts/nvidia-asr-server
uv run server.py

# Point agent-cli at the local endpoint
agent-cli transcribe \
  --asr-provider openai \
  --asr-openai-base-url http://localhost:9898/v1
```

## Testing
- Not run in this branch (CI will cover).

Replace `qwen3:4b` with `gemma3:4b` as default model (#69)

Improve macOS recording notifications (#68)

* Improve macOS recording notifications

* Simplify macOS notification helper

* Simplify more

Improve transcription context sharing (#66)

* Improve transcription context sharing

* Always use raw transcripts for context

Fix run piper script (#64)