Add experimental server with chat completions endpoint

[neilmehta24]

Summary

This PR adds a small authenticated HTTP server for text generation. It exposes health and streaming chat-completion endpoints backed by the existing mlx-engine APIs.

Design

This is not intended to be a general OpenAI-compatible server. Its contract is intentionally limited to the initial text and vision generation path. It assumes one loaded model, a trusted local client, streaming responses, and inline base64 images. Unsupported capabilities fail explicitly rather than being approximated.

generate.py remains the source of truth for generation behavior. The server only validates and translates requests. The loaded model owns the chat template. Clients may supply template variables, but cannot replace the template or override rendering controls. Tool-enabled requests are rejected because the server does not yet emit semantic tool-call events. Images are never fetched from URLs or files.

The continuous batcher previously decoded for up to 500 ms before checking for new requests. This amplified small differences in HTTP request arrival time. It now yields when work is waiting.

Performance

Test	Direct Python	HTTP
Text generation	242.9 tok/s	246.2 tok/s
Vision generation	117.7 tok/s	117.2 tok/s
Uncached vision TTFT	406.0 ms	407.8 ms

Direct Python and HTTP requests used the same loaded model, prompt, and deterministic generation settings. Follow-up parity checks produced identical text, completion-token counts, and stop reasons for both text and vision generation. Qwen2.5 0.5B Instruct 4-bit was used for text generation and Gemma 4 E2B 4-bit for vision generation.