fix(llm): raise on API failure instead of silent empty fallback

[ZwwWayne]

Summary

• AsyncOpenAIWrapper.chat() previously returned a fabricated ChatCompletion with empty content when all retries were exhausted. Callers could not distinguish a real empty response from a total API outage, which silently corrupted downstream results (e.g. evaluation pipelines saw 0-score rounds with no indication of the actual failure).
• Now raises RuntimeError after retries are exhausted so callers can handle failures explicitly.
• Non-retryable exceptions are re-raised immediately instead of being swallowed.
• Replaced print() with logger.warning() for retry diagnostics.
• Cleaned up unused imports (uuid, Choice, ChatCompletionMessage).

Motivation

When using AsyncOpenAIWrapper in a benchmark evaluation harness, intermittent 502/timeout errors from the API resulted in 15 out of 27 tasks silently producing zero-score results — the fallback ChatCompletion(content='') was indistinguishable from a valid but empty model response. The caller had no opportunity to retry at a higher level or log the actual failure.

Test plan

• Verified from lagent.llms.openai import AsyncOpenAIWrapper imports cleanly
• Existing downstream callers that relied on the fallback should add try/except RuntimeError to handle persistent failures gracefully

🤖 Generated with Claude Code