fix(models): close all MultiProvider children after failures - #4438

Summary

This pull request fixes MultiProvider.aclose() so one child provider's ordinary cleanup failure does not prevent later child providers from releasing their resources.

• Closes unique child providers in the existing stable order and preserves identity deduplication.
• Re-raises the exact first ordinary exception after the remaining close attempts finish.
• Continues to propagate direct asyncio.CancelledError immediately.
• Adds regressions for multiple ordinary failures and cancellation after an earlier failure.

Test plan

Passed:

• uv run pytest -q tests/models/test_map.py (17 passed)
• uv run pytest -q tests/models/test_responses_websocket_session.py (11 passed)
• uv run ruff check
• uv run python .github/scripts/check_optional_truthiness.py src/agents
• uv run ruff format and uv run ruff check --fix (900 files unchanged)
• uv run mypy src/agents/models/multi_provider.py

The full Windows test run reached 7,208 passed and 77 skipped, with remaining failures caused by local platform prerequisites unrelated to this change: the Unix tee command is unavailable, symlink creation is not permitted, two UTF-8 examples are decoded with the system GBK code page, and existing realtime Windows tests fail. Full mypy/pyright also report existing errors in untouched sandbox modules; pyright's only error is src/agents/sandbox/util/tar_utils.py:161.

Issue number

N/A

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR