fix: retry pre-response WebSocket server errors

[dk3yyyy]

Summary

This pull request fixes Responses WebSocket server_error frames bypassing provider-suggested model retry policies when they arrive before response processing begins.

It classifies the reported error_type="server_error" and code=None frame as retryable while preserving the existing overload behavior and replay-safety boundaries. Coded non-transient errors remain unaffected, and stateful continuations remain subject to the existing replay-safety gate.

Test plan

• uv run pytest tests/models/test_openai_responses.py -k 'websocket_get_retry_advice' -q (19 passed)
• uv run pytest tests/models/test_openai_responses.py -q (126 passed)
• .agents/skills/code-change-verification/scripts/run.sh (format, lint, full tests, mypy, and pyright passed)
• Independent specification and code-quality reviews passed on the final diff

Issue number

Closes #3990

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 (not applicable; Codex was not used)