fix(models): preserve streamed content part order

[felmonon]

Summary

When a Chat Completions stream emits a refusal before text, the raw stream assigns content_index=0 to the refusal and content_index=1 to the text, but response.completed always assembled text before refusal. Consumers replaying raw events therefore saw a different content layout from the completed response.

This PR assembles the completed assistant message using the content indexes already assigned during streaming. Text-before-refusal behavior stays unchanged, while refusal-before-text now matches the raw event sequence. This is a focused follow-up to the pre-existing mismatch discussed in #3757 and follows the raw-event/final-response alignment required during #3769.

Test plan

• The updated regression test fails on origin/main and passes with this change.
• uv run pytest -q tests/models/test_openai_chatcompletions_stream.py — 57 passed.
• bash .agents/skills/code-change-verification/scripts/run.sh — format, lint, and the full test suite passed.
• Targeted mypy and pyright checks pass for both changed files.
• The repository-wide local typecheck still reports six errors in unchanged current-main files (archive_ops.py, test_any_llm_model.py, and test_run_step_execution.py).

Issue number

Follow-up to #3757; related maintainer direction in #3769.

Checks

• I added regression coverage.
• I ran the full impacted test slice.
• I ran the repository verification script and documented its unrelated typecheck limitation.