Add experimental server with chat completions endpoint (#353)

* Add MLX server

* Simplify MLX server request handling

* Clarify inline image schema names

* Keep chat template rendering model-owned

* Stream HTTP generation on request threads

* Admit batched requests between generation steps

* Handle structured output and assistant prefills

* Validate server request boundaries

* Validate generation requests and prompt usage

* Harden server request and usage accounting

* Format server startup test

* Normalize text messages and close error responses

* Bound stream writes and correct prefill progress

* Harden chat request validation

* Harden server request handling

* Reject decompression bomb images

* Validate complete image requests

* Prevent batched request starvation