#2278 serve: add the legacy /v1/completions endpoint for raw-prompt clients (#2277)

本文

follow-up to #2277 

4th dialect on the route table, thin handler on purpose. the prompt goes in as given, no template, the caller did that one. stream + non-stream, usage included, 0/-1 still mean the rest of the context.

stream_options works with stream:true, include_usage only. the final chunk already sent usage so there was nothing to do.

thinking flags actually do something here now. enable_thinking:false / reasoning_effort:none append the no-think tail from the template, requests without them pass through untouched.

the OSError handler joins the shared exception chain the engine-stuck one sits on, so that part helps all four dialects, not just this endpoint.

5 new tests, docs table line, whole suite green, checked against a live server too.