fix(openai): finalize streams even when usage chunks are omitted

[imMamdouhaboammar]

Summary

Make streamed OpenAI-compatible responses produce an explicit final snapshot even when the endpoint does not return the optional usage chunk requested by stream_options.include_usage.

Root cause

StreamAccumulator used usage !== undefined as the only final-state signal. OpenAI returns usage on a final empty chunk when include_usage is honored, but compatible endpoints can omit usage metadata. In that case every yielded provider snapshot remained marked streaming: true, tying message completion to optional accounting data.

Changes

• track stream completion independently from token usage
• preserve current behavior when a usage chunk is returned
• add finish() so iterator exhaustion can mark the accumulator complete without inventing usage
• make OpenAIModelProvider.stream() emit one final non-streaming snapshot after iterator exhaustion only when no usage chunk already finalized the stream
• retain the existing best-effort incomplete-tool-input fallback on final snapshots
• add integration coverage with a mocked compatible endpoint that streams text and ends without usage metadata

This keeps token accounting when available while making the streaming contract independent of an optional provider feature.