feat(server): stream POST /v1/completions token-by-token (no new nati…

[vaiju1981]

Summary

Stream POST /v1/completions token-by-token — with no new native/JNI code. The streaming
raw-completion path already exists in the JNI layer (requestCompletion/receiveCompletionJson,
exposed as LlamaModel.generate(InferenceParameters) → LlamaIterable), so this is purely Java server
wiring; the earlier "a new requestCompletionStream native method is needed" note was stale.

• OpenAiRequestMapper.toCompletionParameters maps an OpenAI completion request (prompt + sampling)
  to InferenceParameters. The shared sampling/cache/output fields are factored into a new
  applyCommonFields reused by the chat mapper (its tests confirm no behaviour change).
• OpenAiBackend.streamCompletions(request, sink) (default throws UnsupportedOperationException) +
  a LlamaModelBackend impl that drives generate() and emits one OpenAI text_completion chunk per
  token, mapping StopReason → finish_reason (length / stop). A sink IOException (client
  disconnect) cancels the native task via LlamaIterable.close().
• OpenAiSseFormatter.completionChunk builds the chunk; handleCompletions branches on stream:true
  to a new streamCompletions SSE handler that mirrors streamChat (heartbeats, [DONE], graceful
  disconnect).

Test plan

• Affected tests pass locally — new streaming /v1/completions HTTP test (FakeBackend), 16
  mapper tests + 36 server tests + 40 adjacent server tests; Spotless + Javadoc clean.
• CI is green on this branch
• Docs — server class Javadoc updated; TODO.md corrected (stale "new native method" note; marks
  /v1/completions streaming done).

Related issues / PRs

Server follow-up from the OpenAI-compatible endpoint work.

Note for reviewer

Remaining consumers are queued as follow-ups (same generate()-driven pattern, no new infrastructure):
token-streaming Ollama /api/generate and Continue's native POST /completion. Verification gap: the
streaming path is proven by the FakeBackend HTTP test (SSE framing, [DONE], the mapper) but not yet
by a real-model streaming test — the gated OpenAiServerCompletionIntegrationTest is the place to add one.

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes (if there are, I have notified the maintainer privately per SECURITY.md)