Return http status code for ApiError

[max-zheng]

Preserves the HTTP StatusCode from non-2xx responses by changing OpenAIError::ApiError to wrap a new ApiErrorResponse { status_code, api_error } struct. This lets callers distinguish 400 / 404 / 5xx without pattern-matching on ApiError.r#type strings — the original motivation in #549, e.g. for gateways that proxy to non-OpenAI-compatible upstreams.

Implements option (2) from the issue discussion (the maintainer's preferred shape — wrapping in a struct rather than a struct-variant, so a single value can be passed/extended later).

This is a breaking change for callers that pattern-match on OpenAIError::ApiError(_): the inner type is now ApiErrorResponse rather than ApiError, accessed via .api_error. Per the issue thread, a dedicated minor release was suggested.

Closes #549.

Changes

• error.rs: add ApiErrorResponse (status + api_error, with Display/Error impls); change variant to ApiError(ApiErrorResponse). The existing ApiError struct and its JSON-body Display are untouched.
• client.rs read_response: wrap both 5xx and non-2xx construction sites with the captured status.
• middleware/retry/openai.rs: capture status_code before consuming the response body in the 429 path; wrap both the transient and insufficient_quota error returns.

[64bit]

Thank you for the PR!