fix(openai): surface failed response streams

[EfeDurmaz16]

What

Raise ProviderResponseError when the OpenAI Responses API emits response.failed. Preserve the provider error message, code, body, and retryability signal.

Why

response.failed is a terminal provider failure, but the adapter currently continues and emits a normal StreamEnd. Applications can therefore persist partial output as a successful completion.

How

• Convert the official response error payload into ProviderResponseError.
• Mark rate_limit_exceeded, server_error, and vector_store_timeout as retryable.
• Keep already streamed partial output available to the caller while raising.

OpenAI event reference: https://platform.openai.com/docs/api-reference/responses-streaming/response/failed

Verification

• uv run --frozen pytest -q: 594 passed
• Ruff format and lint passed
• mypy passed
• ty passed

Suggested release label: fix.

[vercel]

@EfeDurmaz16 is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.