fix(chatgpt): preserve non-success response errors

[gold-silver-copper]

Summary

• check ChatGPT Responses non-streaming HTTP status before parsing the SSE body
• return CompletionError::from_http_response(status, body) for non-2xx responses so status/body helpers work
• add unit coverage for 401 and 429 status/body propagation
• add live-recorded ChatGPT cassette coverage for the 401 Responses API error path
• document the provider checklist requirement for non-2xx completion errors

Fixes #2037

Tests

• cargo fmt
• RIG_PROVIDER_TEST_MODE=record CHATGPT_ACCESS_TOKEN=invalid-token-for-recording CHATGPT_ACCOUNT_ID=invalid-account cargo test -p rig --all-features --test chatgpt chatgpt::cassette::http_errors::nonstreaming_unauthorized_preserves_status_and_body -- --nocapture --test-threads=1
• cargo test -p rig-core providers::chatgpt::tests::completion_http_non_success_preserves_status_and_body -- --nocapture
• cargo test -p rig --all-features --test chatgpt chatgpt::cassette::http_errors -- --nocapture --test-threads=1
• cargo clippy --all-targets --all-features
• cargo test -- --test-threads=1

Cassette notes: the committed ChatGPT cassette was recorded live against https://chatgpt.com/backend-api/codex using an intentionally invalid access token to produce a deterministic 401. I did not live-record a 429 fixture because reliably forcing ChatGPT rate limiting would require intentionally exhausting a real account/rate limit; the 429 path is covered by the unit test instead.