feat(scripts): add MLX-LM server for fast Apple Silicon inference

[basnijholt]

Summary

• Add run-mlx-lm.sh script that starts an OpenAI-compatible MLX LLM server on port 8080
• Default model: mlx-community/Qwen3-4B-4bit (configurable via MLX_MODEL env var)
• Only works on Apple Silicon (M1/M2/M3/M4) - provides significantly faster inference than Ollama
• Add MLX-LLM pane to start-all-services.sh Zellij layout
• Update README with MLX-LM documentation

Usage

# Start the server
./scripts/run-mlx-lm.sh

# Or with custom model
MLX_MODEL=mlx-community/Qwen3-8B-4bit ./scripts/run-mlx-lm.sh

# Use with agent-cli
agent-cli transcribe --llm --llm-provider openai --openai-base-url http://localhost:8080/v1
agent-cli autocorrect --llm-provider openai --openai-base-url http://localhost:8080/v1

Test plan

• Run ./scripts/run-mlx-lm.sh on Apple Silicon Mac
• Verify server starts and responds to requests
• Test with agent-cli transcribe --llm --llm-provider openai --openai-base-url http://localhost:8080/v1