feat(llm): add local MLX LLM provider for fully offline voice input

[erning]

Summary

• Add local MLX LLM provider for on-device text correction on Apple Silicon, enabling Koe to run fully offline — both ASR and LLM correction — with no network access or API keys required
• Add mode field to model manifests to distinguish ASR vs LLM models, with mode-based filtering in the Setup Wizard
• Add MLX as a selectable LLM provider in the Setup Wizard with model picker, download/status controls

Motivation

Privacy-conscious users and those in restricted network environments can now use Koe entirely on-device. By pairing local ASR (MLX Qwen3-ASR or Apple Speech) with local LLM correction (MLX Qwen3), all voice data stays on the machine — nothing is sent to any server.

GPU memory for fully offline mode (MLX ASR + MLX LLM):

• Lightest (ASR 0.6B + LLM 0.6B): ~1.2 GB — runs comfortably on any Apple Silicon Mac
• Heaviest (ASR 1.7B + LLM 1.7B): ~2.9 GB — still fits easily in 8 GB unified memory

Performance: The 4-bit quantized Qwen3 models generate at hundreds of tokens per second on Apple Silicon via MLX. A typical voice input correction completes in under a second with the 0.6B model.

Changes

Model manifests:

• Add Qwen3-0.6B-4bit and Qwen3-1.7B-4bit LLM model manifests
• Add mode field ("asr" / "llm") to all MLX manifests for filtering
• Add mode to ModelManifest struct and FFI JSON output

Setup Wizard:

• ASR pane: filter model popup by mode == "asr"
• LLM pane: add provider popup (OpenAI Compatible / MLX) with independent model selection, download, and status controls for MLX

Runtime:

• Swift MLXLlmManager: model loading, text generation with thinking token stripping, KV cache cleanup
• Swift C bridge: koe_mlx_llm_generate, koe_mlx_llm_free_string, koe_mlx_llm_unload_model
• Rust MlxLlmProvider: implements LlmProvider trait via C FFI to Swift, with timeout support
• LlmProvider trait migrated to #[async_trait] for dynamic dispatch (Box<dyn LlmProvider>)
• run_session() dispatches to MLX or OpenAI provider based on llm.provider config
• llm_is_ready() and warmup logic updated for MLX (no HTTP needed)
• Skip interim history for MLX (small models don't benefit, reduces inference time)

Config:

• New fields: llm.provider ("openai" default, or "mlx"), llm.mlx.model
• Default config template updated

Test plan

• Build: make build succeeds
• LLM pane: selecting OpenAI shows API fields; selecting MLX shows model picker
• ASR pane: only ASR models shown (no LLM models mixed in)
• Set llm.provider: mlx, download model, dictate — local correction works
• Set llm.provider: openai — existing cloud LLM flow unchanged
• Fully offline: set both ASR and LLM to local providers, disable network — works end-to-end