feat(backends): add Moonshine streaming STT backend

[Theohox]

Adds a new 'moonshine' recipe for Moonshine speech-to-text using the moonshine_voice C++ streaming API. Integrates with existing OpenAI-compatible /v1/audio/transcriptions and WebSocket /realtime endpoints.

Includes:

• MoonshineServer backend (HTTP /inference, TCP streaming)
• IStreamingTranscriptionServer interface for generic streaming backends
• TcpJsonlClient for line-delimited JSON over TCP
• RealtimeSessionManager integration for WebSocket/audio forwarding
• ModelManager support for moonshine cache resolution and download
• CPU-only, cross-platform (Linux/Windows x86_64)
• End-to-end smoke test

The existing Whisper path is unchanged.

Requesting @bitgamma to review per discord comments

[bitgamma]

Thanks for your PR!

I see the current implementation is relying on system-wide Python installation, which is not something we can rely on. Generally, we try to avoid Python-based backends for this reason. The only Python-based backend we have is vLLM but it has been packaged to be self-contained and not rely (or interfere with) the system Python. Similar work must be done before this can be considered viable.

Additionally, I see models are not being downloaded from huggingface. Adding another download repository is something we also want to avoid. I'd much prefer if you could find or re-upload the models on HF instead. I also see voices are being downloaded in the user folder, outside the usual lemonade model folder - this is another thing we want to avoid.

In general, this needs to be reworked to follow the same patterns as the other backends.

[Theohox]

Thank you!