fix: add SSE streaming support to MLX engine

[Turcy76]

Problem

When using mlx_engine.py as the backend (Option B), agent.py shows no response for chat messages.

The root cause is that agent.py's stream_llm() sends "stream": true and expects SSE (Server-Sent Events) format, but mlx_engine.py only returns a single JSON response, ignoring the stream parameter.

Additionally, the model detection shows auto ? because /props returns lowercase "9b" while agent.py checks for uppercase "9B".

Changes

• generate_stream() — New streaming generator using mlx_lm.stream_generate(), with filtering for <think> tags and special tokens
• _handle_chat_stream() — SSE response handler that sends data: {...}\n\n chunks followed by data: [DONE]
• _handle_chat() — Now checks stream parameter and routes to streaming or normal handler
• clean_response() — Extracted shared cleanup logic (strip special tokens + thinking tags)
• /props endpoint — Fixed model_alias to use uppercase (9B/35B) to match agent.py's detection

Testing

Tested locally with MLX + Qwen3.5-9B-MLX-4bit on Apple Silicon M4:

• ✅ agent.py now displays streamed responses in both auto and raw modes
• ✅ Intent classification (non-streaming) still works
• ✅ Model detection shows auto 9b instead of auto ?
• ✅ /bench command works correctly